Multi-Agent Belief

Deutsches
Forschungszentrum
Research
fOr KOnstliche
Intelligenz GmbH
Report
RR-93-29
Representing Belief in Multi-Agent Worlds

via
Terminological Logics
Armin Laux
September 1993
Deutsches Forschungszentrum fur Kunstliche Intelligenz

GmbH
Postfach 20 80 Stuhlsatzenhausweg 3
67608 Kaiserslautem, FRG 66123 Saarbrticken, FRG
Tel.: (+49631) 205-3211/13 Tel.: (+49681) 302-5252
Fax: (+49631) 205-3210 Fax: (+49681) 302-5341
Deutsches Forschungszentrum
fOr
KOnstliche Intelligenz
The German Research Center for Artificial Intelligence (Deutsches Forschungszentrum fUr KOnstliche
Intelligenz, DFKI) with sites in Kaiserslautern and SaarbrOcken is a non-profit organization which was
founded in 1988. The shareholder companies are Atlas Elektronik, Daimler-Benz, Fraunhofer
Gesellschaft, GMD, IBM, Insiders, Mannesmann-Kienzle, SEMA Group, and Siemens . Research
projects conducted at the DFKI are funded by the German Ministry for Research and Technology, by
the shareholder companies , or by other industrial contracts.
The DFKI conducts application-oriented basic research in the field of artificial intelligence and other
related subfields of computer science . The overall goal is to construct systems with technical
knowledge and common sense which - by using AI methods - implement a problem solution for a
selected application area . Currently, there are the following research areas at the DFKI :
o Intelligent Engineering Systems

o Intelligent User Interfaces
o Computer Linguistics
o Programming Systems
o Deduction and Multiagent Systems
o Document Analysis and Office Automation .
The DFKI strives at making its research results available to the scientific community . There exist many
contacts to domestic and foreign research institutions, both in academy and industry . The DFKI hosts
technology transfer workshops for shareholders and other interested groups in order to inform about
the current state of research.
From its beginning, the DFKI has provided an attractive working environment for AI researchers from
Germany and from all over the world. The goal is to have a staff of about 100 researchers at the end
of the building-up phase .
Friedrich J. Wendl
Director
Representing Belief in Multi-Agent Worlds via Terminological
Logics
Armin Laux
DFKI-RR-93-29
This work has been supported by a grant from The Federal Ministry for Research
and Technology (FKZ IWT-9201).
Deutsches Forschungszentrum fUr Kunstliche Intelligenz 1993

This work may not be copied or reproduced in whole or in part for any commercial purpose. Permission to
copy in whole or in part without payment of fee is granted for nonprofit educational and research purposes
provided that all such whole or partial copies include the following : a notice that such copy ing is by perm ission of
Deutsches Forschungszentrum fUr Kunstliche Intelligenz, Kaiserslautern , Federal Republic of Germany ; an
acknowledgement of the authors and individual contributors to the work; all applicable portions of this copyright
notice. Copying, reproducing, or republishing for any other purpose shall require a licence with payment of fee to
Deutsches Forschungszentrum fUr Kunstliche Intelligenz.
Representing Belief in Multi-Agent Worlds via
Terminological Logics
Armin Laux
German Research Center for Artificial Intelligence (DFKI)
Stuhlsatzenhausweg 3
66123 Saarbriicken, Germany
e-mail: laux@dfki.uni-sb.de
Abstract
In multi-agent systems a group of autonomous intelligent systems, called
agents, acts and cooperates in a world in order to achieve certain goals. Such
systems are in general assumed to have no central control structure and hence
each agent can only perform actions that are based on his local knowledge and
on his local beliefs . In the literature knowledge of agents is mostly represented
under the view that knowledge is true belief. On the other hand, if agents are
acting in a (real) world their knowledge often is obtained by perception and
communication, and hence typically is not true. Thus, the use of belief-where
agents may have false beliefs-seems more appropriate than the use of knowledge
in multi-agent systems.
Terminological logics provide a well-investigated and decidable fragment of
first-order logics that is much more expressive than propositional logic and well
suited to describe a world agents are acting in. However, knowledge or belief of
agents can only be represented in a very limited way. In this paper we investigate
how terminological logics can be extended in such a way that belief of agents
can be represented in an adequate manner. We therefore exemplarily extend
the concept language ALC by a modal operator 0, which is indexed by agents.
Thereby, 0iCP represents the fact "agent i believes cp". This belief operator will be
interpreted in terms of possible worlds using the well-known modal logic KD45.
This extended language ALCa provides a uniform formalism to describe both,
a world agents are acting in and the beliefs agents have about this world and
about their own and other agents' beliefs. Thus, it can be seen as a two-
dimensional extension of ALC which allows both, reasoning about objective facts
that hold in the world and reasoning on the level of possible worlds. We will give
sound and complete algorithms to check consistency of the represented beliefs
and to decide whether an ALCa-sentence is logically entailed by the beliefs of
agents. Hence, when acting in a world agents can use beliefs which are explicitly
represented as well as implicit beliefs that are entailed by their knowledge base.
Contents
1 Introduction 3
2 Syntax and Semantics of ACB 5

2.1 The Concept Language AC . 5
2.2 The Extended Language ACB 7
3 Testing Satisfiability of ACB-formulas 9

3.1 The ACB Frame Algorithm . . .. . . 9
4 Optimization and Computing ACB-Inferences 20

4.1 Computing AC test sets 21
4.2 Restricted AC- TBoxes . 28
4.3 Computing ACB-Inferences 30
5 Conclusion 31
2
1 Introduction
Research on the field of multi-agent systems deals with the question how a group of
autonomous intelligent syst ems , called agents, can cooperate in order to achieve certain
goals (see, e.g., [6, 15]). As an example, a forwarding agent a and a shipping agent b
may cooperate in order to carry out overseas transportation orders.
Although the tasks that multi-agent systems are required to perform are normally
stated in terms of the global behavior of the system, the actions that an agent performs
can depend only on his local knowledge and on his local beliefs. Thus, there is a close
relationship between knowledge, belief, and action in multi-agent systems (see, e.g.,
[27,28, 16]). Suppose, in the above example agent a wants to offer a price for carrying
out some transportation order 01' If he believes that there is no other forwarding
agent who also can carry out 01, he will most likely offer another price as in the case
where he believes that there is a competitor for this order. And if he even knows
that there is no competitor for this order, he perhaps offers an exorbitant price. In a
recent paper [21] we investigated how knowledge of agents can be represented on the
basis of terminological logics, whereby we used the classical view of knowledge as true
belief. That means, an agent knows <.p if he believes that <.p holds and <.p actual does
hold. On the other hand, as pointed out in, say [23], the knowledge represented in a
knowledge base typically is not true. Thus, the use of beliefs-where agents may have
false beliefs- seems more appropriate than knowledge for formalizing the reasoning and
deduction of a knowledge base. In the current paper yve concern with the question how
agents can be equipped with beliefs about the world they are acting in, about beliefs
of other agents, and also about their own beliefs. Thereby, it should be taken into
consideration that different agents may have different beliefs about the same notions.
For example, forwarding agent a may believe that company XY is a rich company and
a good client, while forwarding agent B believes that company XY is rich but not a
good client. Using the language ALes, which is presented in the next section, this can
be formalized by
Da( company XY: rich-company n good-client) and
D b ( company XY: rich-company n -,good-client)
respectively, where Di<.p is to be read as "agent i believes <.p"

Since the work of Hintikka [18], modal logics have widely been accepted to be an
adequate formalism for representing knowledge and belief of agents. The intuitive
idea here is that besides the real world agents can imagine a number of other worlds
(situations) to be possible. By imposing various conditions on this possibility relation,
we can capture a number of interesting axioms. For example, if we require that the
world that the agent finds himself in is always one of the worlds he considers possible
(which amounts to saying that the possibility relation is reflexive), then it follows that
the agent does not know false facts. When using a possibility relation which captures
axioms of knowledge (belief) an agent is said to know (believe) a fact <.p iff <.p is true
3
in all worlds he thinks to be possible. For example, an agent knows (believes) that
there exists a monster of Loch Ness if there is such a monster in all worlds he considers
possible. To express the beliefs of an agent a in this approach a binary operator
BELIEF (a, r.p) is used, where r.p is a formula over some logical language C. If we
want to devise a formalism for representing the beliefs of agents we have to take two
decisions. Firstly, we have to decide what the general properties of belief are we want
this formalism to capture. Secondly, we have to choose a suitable logical representation
language C which allows to describe the beliefs of agents.
There are many approaches to determine axioms characterizing belief (see, e.g.,
[22, 28, 24, 25, 26, 12, 17]). We will use the following axiomatization which has been
most commonly used in the literature. The first of these properties states that an
agent does not believe false facts. That means, an agent cannot believe both a fact
and its negation, though he can believe facts which actually do not hold in the world.
Secondly, if an agent believes a fact then he believes that he believes it (positive
introspection), and if he does not believe in a fact then he believes that he does not
believe in it (negative introspection). From this it follows, e.g., that agents believe that
their beliefs are true (weak reflexivity). Finally, the probably most important property
is that agents can reason on the basis of their beliefs. For example, suppose agent a
believes that each truck which is owned by John can be used to transport gasoline and
he believes that John owns the truck truck-l. In this case, agent a must be able to
conclude that John's truck truck-l can (probably) be used to transport gasoline, and
thus may negotiate with John for a transportation order.
As logical language to describe belief of agents we will use a terminological logic.
Terminological logics provide a well-investigated and decidable fragment of first-order
logics that is much more expressive than propositional logic. They are based on the
work of Brachman and Schmolze [9] and have been developed as a structured formalism
to describe the relevant concepts of a problem domain and the interactions between
these concepts. Starting with atomic concepts (unary predicates) and roles (binary
predicates), one therefore defines complex concepts with the help of operators provided
by a concept language, and interactions between (complex) concepts are expressed by
a set of so-called terminological axioms. On the other hand, by so-called assertional
axioms, objects can be associated with concepts and relationships between objects can
be defined via roles. For example, we can use these logics to represent facts like "each
truck which is owned by John can be used to transport gasoline" or "John owns truck-l
which is a truck" .
In the literature, a lot of concept languages have been considered (see, e.g., [8, 29,
3]). But they all have in common that they are only suitable for representing objective
facts about the world, and knowledge or beliefs of agents can only be represented
in a very limited way. Thus, we need an extended concept language which allows
the representation of belief according to the above given (informal) axiomatization.
Since the work of Schild [31] it is known that the concept language ACe provides a
4
terminological logic which is a notational variant of the propositional modal logic K(m).
However, it is not investigated there how to extend this logic to a two-dimensional logic
which allows reasoning on both the objective level and the level of possible worlds. In
order to combine both levels one has to define syntax and semantics of an extended
language. Baader and Ohlbach [5] present a multi-dimensional extension of ACe,
where multi-modal operators can be used at all levels of the concept terms and they
can be used to modify both concepts and roles. However, the underlying logic is simply
the basic modal logic K, and it is not yet clear how to extend their approach in such
a way that modal logics different from K can be handled. Moreover, they could not
succeed in proving completeness of their satisfiability algorithm.
In this paper we will present a different extended language where (sequences of)
modal operators are only allowed in front of terminological and assertional axioms. This
language allows one to interpret the modal operators w.r.t. modal logics different from
K, e.g., S4 (see [21]) or KD45 (in the present paper). This language, called ACes, can be
seen as a two-dimensional representation language with terminological and assertional
axioms as primitives where each primitive may describe a part of the world and each
agent can believe a set of such primitives to hold in the world. The modal operators,
which are indexed with agents, are interpreted in terms of possible worlds in such a way
that they satisfy the above axiomatization of belief, what amounts in using the modal
logic KD45. Thus, the resulting language provides a uniform formalism to describe
both, a world agents are acting in as well as the beliefs agents have about this world and
about their own and other agents' beliefs. We will give sound and complete algorithms
for deciding satisfiability of ACes-formulas and for testing whether an ACes-formula
is entailed by a given set of ACes-formulas. Hence, when acting in a world agents can
use beliefs which are explicitly represented as well as implicit beliefs that are entailed
by their knowledge base.
2 Syntax and Semantics of ALeE
In this section we will formally introduce the language ACes which extends the concept
language ACe by a modal operator 0i for each agent. Syntax and semantics of ACe
and ACCB are given in Subsections 2.1 and 2.2, respectively.
2.1 The Concept Language ALe
Terminological logics provide two formalisms to describe a problem domain: a termino-

logical formalism to represent taxonomical knowledge by defining concepts, which can
be seen as sets of objects, and an assertional formalism which can be used to describe
concrete objects. Therefore, one starts with a set of atomic concepts (unary predicates)
and a set of roles (binary predicates). In the concept language ACe concepts are then
5
built up from atomic concepts, the top concept T, the bottom concept 1-, and roles
inductively by:
1. Each atomic concept, T, and 1- are concepts.
2. If e and D are concepts and R is a role, then
(a) enD (concept conjunction),

(b) eUD (concept disjunction),
(c) -.e (concept negation),
(d) V R. C (value restriction), and
(e) 3 R. C (exists restriction)
are concepts.
An interpretation I is a function over some non-empty domain L\ I which maps each

atomic concept e to a subset e I of L\ I, each role R to a subset RI of L\ I X L\ I, T to
L\I, and 1- to 0. Furthermore, n is interpreted as set intersection, U as set union, and
-. as set complement W.r.t. L\I. The value and the exists restrictions are interpreted by
[VR.CjI = {d E L\I I Vd': (d,d') E RI -+ d' Eel}

[3R.CjI = {d E L\I I 3d' : (d,d') E RI i\ d' Eel}
For example, if man and truck are atomic concepts and owns is a role we can define
the concept of men who own a truck by man n 3 owns. truck.
The taxonomical knowledge of a problem domain can be defined by an AC- TBox
(terminology), which consists of a finite set of terminological axioms. A terminological
axiom is of the form
e= D (concept equivalence) or
e i- D (negated concept equivalence)

where e, D are concepts. An interpretation I satisfies e = D iff e I = DI and it
satisfies e i- D iff e I i- DI. An interpretation I satisfies an AC- TBox T iff I
satisfies each axiom in T. For example, if carrier, person, and truck are concepts and
owns is a role, we can define exactly the persons who own a truck to be a carrier by
carrier = person n 3 owns. truck.
The assertional formalism of AC allows one to introduce concrete objects by stat-

ing that they are instances of concepts and roles: If a is an object and ea concept,
then a : e is a concept instance. If a and b are objects and R is a role, then aRb is
6
a role instance. Concept instances and role instances are called assertional axioms,
and a finite set of assertional axioms is called an A.cC-ABox. An interpretation I
maps objects to elements of its domain 6,1 and satisfies a : C iff a I E CI, and aRb iff
(aI, bI ) E RI. We assume that different objects in an ABox are mapped to different
elements in 6,1 (unique name assumption). An interpretation I satisfies an A.cC-ABox
A iff I satisfies each axiom in A. As an example, if John and truck-l are objects, we
can express that John owns truck-l which is a truck by the assertional axioms
John owns truck-l and truck-l : truck.
Thus, we can describe the relevant concepts of a problem domain by terminological

axioms, i.e., by an A.cC-TBox, and properties of objects as well as relations between
them by assertional axioms, i.e., by an A.cC-ABox. We sayan interpretation I satisfies
a set AXI, ... , AXn of terminological and assertional axioms iff I satisfies each of these
axioms. We then write I 1= AXI,"" Axn .
For sake of simplicity we will sometimes use the expressions C ~ D and C g D
where C and D are concepts. An interpretation I satisfies C ~ D iff C I ~ DI and
it satisfies C (l D iff C I C1 DI. The next lemma states that these expressions are
abbreviations for certain terminological axioms.
LelTIlTIa 2.1 Let C and D be concepts, and let I be an interpretation. Then
1. I satisfies C ~ D iff I satisfies -,C U D = T.
2. I satisfies C (l D iff I satisfies -,C U D -::J T.
Proof: For 1., firstly suppose I satisfies C ~ D. Then for each element d in 6,1 either
dE [-,CF or both d E C I and dEDI holds. That means, I satisfies -,CU(CnD) =T
what can be simplified to -,C U D = T. Conversely, suppose I satisfies -,C U D = T.
Then for each element d E 6,1 either d rf. C I or dEDI holds. Thus, from d E C I
follows dEDI, i.e., C I ~ DI. The proof of 2. is analogous. 0
For example, if truck and vehicle are concepts we can define each truck to be a
vehicle by truck ~ vehicle, what is an abbreviation for -,truck U vehicle = T.
2.2 The Extended Language ALes
Now we will introduce the language A.cCB which extends A.cC by a new operator OJ
for each agent i.1 We allow these operators in front of terminological and assertional
1 In the following, we will abbreviate agents by numbers, and we suppose only a finite number of
agents to be given .
7
axioms. Thereby, the operator OJ, read as "agent i believes", allows us to express the
beliefs agent i has about the world, about beliefs of other agents, and about his own
beliefs. We extend the definition of terminological and assertional axioms as follows.
If TAis a terminological axiom, then OJ TA and --.ojTA are terminological axioms

as well.
If CJ is a concept instance, then OJ CJ and --. OJ CJ are concept instances as well.
If RI is a role instance, then OJ RI is a role instance as well.
Note, that we do not allow formulas of the form --.OJ(aRb). The reason for this re-
striction is that such axioms would be equivalent to stating that there exists a world
in which the role instance aRb does not hold. And negation of roles is not allowed in
AL:C.
These extended assertional and terminological axioms are called AL:CB-formulas
and can, e.g., be used to state that agent i believes that each truck is a vehicle by
OJ (truck ~ vehicle).
Analogously, the AL:CB-formulas OJ--.oj (vehicle-l : truck) and oj--.oj (vehicle-l :
truck) are to be read as "agent i believes that agent j doesn 't believe that vehicle-l is a
truck" and "agent i believes that he doesn't believe truck-l to be a truck" , respectively.
Allowing OJ immediately in front of concepts (possibly ojG may be interpreted as "the
set of individuals agent i believes to be a C") causes essential algorithmic problems
and is out of the scope of this paper.
We will interpret the operators OJ in terms of possible worlds, i.e. , besides the real
world there exist a number of worlds agents consider to be possible. If agent i considers
world w' possible at world w , we say w' is accessible from w by agent i. The accessibility
relation of agent i is given by all pairs (w , w') such that w' is accessible from w by agent
i. Since different worlds are possible in our approach, the interpretation of concepts
and roles in AL:CB-formulas depends on the world we are currently speaking of. That
means, in different worlds concepts may contain different objects and roles may contain
different pairs of objects. This will be expressed by taking an additional parameter, the
world parameter, into consideration when interpreting concepts and roles. Formally,
we use the notion of a J( -interpretation J(I which consists of a non-empty domain 6"KI
and maps objects to elements in 6"KI while satisfying the unique name assumption,
atomic concepts to subsets of 6"KI x W, T to 6"KI X W, J.. to 0 x W, and roles to
subsets of 6"KI x 6"KI X W. Furthermore, n is interpreted as set intersection, U as set
union, and --. as set complement w.r.t. 6"KI X W, and the value and exists restrictions
are interpreted by
[\I R.C]KI = {(d,w) I (d' , w) E CKI for each d' with (d , d',w) E RKI}
[3R.C]KI = {(d,w) I (d',w) E CKI for some d' with (d , d' , w) E RKI}.
8
Definition 2.2 A /(ripke structure /( is a triple (W, r, /([). Thereby, W is a non-
empty set of worlds, r is a finite set of accessibility relations, one accessibility relation
/ i for each agent i, and /([ is a /( -interpretation.
The satisfiability of an ALCs -formula F in a /(ripke structure /( = (W, r,I<[) and a

world w E W , written as /(, w F= F, is recursively defined by:
/(, w F= C = D iff {d I (d,w) E CKI} = {d I (d,w) E DKI}

/(, w 1= C =f D iff {d I (d,w) E CKI} =f {d I (d,w) E DKI}
/(, w 1= a: C iff (a ,w)ECKI
/(, w F= aRb iff (a , b, w) E RKI
/(, w F= DiG iff /(, w' F= G for each world w' with (w, w') E /i
/(, w 1= 'DiG iff there is a world w' with (w, w') E /i and /(, w' ~ G
where G is an ALCs-formula, C , D are concepts, a , b are objects, and R is a role.

A set F1 , . . . , Fn of ALCs -formulas is sat~sfiable iff there exists a Kripke structure
/( = (W, r, /([) and a world Wo E W such that /(, Wo F= Fi for i = 1, .. . , n. We then
wri te /( F= F1 , .. , Fn .
In the following we will use the notion modality to denote (negated) indexed 0
op erators, and ALCs-formulas without any modalities are called ALC-formulas. For
example, the ACs-formula o;,oj(vehicle-l : truck) contains the modalities 0; and
, 0 j, and the ALCs-formula vehicle-l : truck is an ALC-formula.
3 Testing Satisfiability of ALes-formulas

Using ALCs-formulas, a "real world" and belief of agents can be defined as follows. The
real world is given by a finite set of AC-formulas, and the belief of agent i is given by
a finite set of ALCs-formulas with the leading modality oi. Of course, we do not only
want to represent a world and beliefs of agents, but we are interested in algorithms to
test (i) consistency of the represented facts, i.e., whether a given set of ALCs-formulas
is satisfiable, and (ii) whether an ALCs-formula is a logical consequence of a given set
of ALCs-formulas. In this section we will give an algorithm for testing satisfiability
of a set of ALCs-formulas. Building upon this we will show how to decide whether or
not an ALCs-formula is a logical consequence from a given set of ALCs-formulas in
Section 4.
3.1 The A.ces Frame Algorithm
We will now present an algorithm for testing satisfiability of a finite set F1 , ... , Fn of
ALCs-formulas. By definition, a set F1 , . ,Fn of ALCs-formulas is satisfiable iff there
9
exists a Kripke structure K such that K 1= Fl, ... ,Fn. Of course, we are not interest-
ed in arbitrary Kripke structures to satisfy F l , ... ,Fn , but only in Kripke structures
which interpret the belief operators 0 in Ft, ... ,Fn in such a way that they satisfy the
properties described in Section 1. We therefore introduce the notion of KD45 Kripke
structures.
Definition 3.1 A set Ft, .. . ,Fn of AC f3 -formulas is KD45-satisfiable iff there ex-
ists a K ripke structure K = (W, r, K I) which satisfies Fl, ... ,Fn and which has the
properties
(Pi) if K, W F= DiF then K, w F= -,Di-,F
(P2) if K, W F DiF then K, w F DjDjF
(P3) if K, W F -,DjF then K, w F Dj-,DjF
for each AC f3 -formula F , for each agent i, and for each world w E W. 2 A Kripke
structure which satisfies (Pi), (P2) , and (P3) is called KD45 Kripke structure.
Property (PI) corresponds to "an agent cannot believe in both a fact and its nega-
tion", (P2) to "if an agent believes something, then he believes that he believes it",
and (P3) to "if an agent does not believe in a fact then he believes that he does not
believe in this fact". The property "agents must be able to reason on the basis of their
beliefs", is guaranteed by choosing Kripke structures for the representation of belief
(d., e.g., [17]).
It is a well-known fact that K = (W,r,KI) is a KD45 Kripke structure if the
accessibility relation ,j
of each agent i is serial, Euclidean, and transitive (see, e.g.,
[25]). A relation , ~ W x W is
serial iff for each u in W there is a v in W such that (u, v) E "
Euclideaniffforallu,v,win Wholds: if(u,v) E ,and (u,w) E ,then (v,w) E 1,
transitive iff for all u, v, win W holds: if (u, v) E , and (v, w) E , then (u, w) E ,.
We will use the standard notation 0 j F as an abbreviation for -,Dj-,F such that the
properties (PI) and (P3) can be rewritten as
(PI') if K, w 1= DjF then K, w 1= 0 F j
(P3') if K, w 1= 0 F then K, w F D 0
j j j F.
In the following we will use the one or the other version of these two properties, whichev-
er is more appropriate.
2Since these properties hold for arbitrary worlds this amounts in saying that all these properties are
mutually believed, i.e., each agent's belief has these properties, each agent believes that each agent's
belief has these properties and so on .
10
To keep notation simple we transform ACB-formulas into negation normal form.
An ACB-formula (concept) is in negation normal form iff in the formula (concept)
negation signs occur immediately in front of atomic concepts only. Concepts can be
transformed into an equivalent negation normal form by the rules
,(C n D) -t ,C U ,D ,(\I R. C) -t :3 R. ,C
,(CUD) -t ,Cn,D ,(:3 R. C) -t \I R., C
where C is a concept and R is a role (see, e.g., [20]). Building upon this, ACB-formulas
can be transformed into negation normal form by applying the rules
"F -t F ,(C = D) -t an: (C n ,D) U (,C n D)

,DiF -t Oi,F ,(Ci-D) -t C=D
,OiF -t Di,F ,( a : C) -t a: ,C
to the outermost negation sign. Thereby, F is an ACB-formula, C, D are concepts,

a is an object, and an is a new object. For example, the negation normal form of the
AC B- formula
IS
where an is a new object . The next lemma states that an ACB-formula is KD45-
satisfiable iff its negation normal form is KD45-satis~able.
Lemma 3.2 Let F be an ACB-formula, F* be the negation normal form of F, and

I< be a KD45 Kripke structure. Then I< F= F iff there is a KD45 Kripke structure K'
such that I<' F= F* .
Proof If we apply one of the rules " F - t F, ,DiF - t Oi,F, ,OiF - t Di,F, or
,(a: C) - t a : ,C to an ACB-formula F, then I< obviously satisfies the formula on
the left hand side of the rule iff I< satisfies the right hand side of the rule.
Now suppose I< = (W,f,I<I) and I< 1= ,(C = D), i.e., I<,wo F= ,(C = D) for
some world Wo in W. In this case there is an element u E 6 KJ such that (u, wo) is either
in CKJ and in [,DlKJ or in [,ClKJ and in DKJ. Let now K' = (W', fl, I<r) be a Kripke
structure which is identical to K but a~J := u. Then, I<' is a KD45 Kripke structure3
and I<', Wo F= an : (C n ,D) U (,C n D). Conversely, if I<' is a KD45 Kripke structure
such that I<',wo F= an: (C n ,D) U (,C n D), then obviously I<',wo F= ,(C = D). 0
If F is an ACB-formula in negation normal form it has a (possibly empty) leading

sequence 0* = Oil' .. aim of non-negated modalities where each Oij is either D or 0 and
3Note, that an is a new element and the unique name assumption only has to hold for objects
occurring in an ABox.
11
each index i j is an agent. We now replace each subsequence of modalities indexed with
the same agent in o by the last modality in this subsequence. The obtained ACCs -
formula is called the KD45 normal form F' of F. For example, the KD45 normal form
of DIOI02D2Dl(a : C) is given by OlD 2 D 1(a : C). As an immediate consequence of
Proposition 4.27 in [10], for each KD45 Kripke structure I< = (W,r,I<I) and for each
world w E W holds that I<, w 1= F iff I<, w 1= F'.
Assumption: In the following we suppose each ACCs-formula to be in KD45
normal form (and thus especially in negation normal form).
To formulate a calculus for testing KD45 satisfiability of ACCs formulas we in-
troduce the notions of labeled ACCs-formulas and of a world constraint system. A
labeled ACCs-formula consists of an ACCs-formula F together with a label w, written
as F II w. Thereby, w is a constant representing a world in which F holds. A world
constraint is either a labeled ACCs-formula or a term w ~i Wi. The constants wand
Wi represent worlds and ~i represents the accessibility relation of agent i. A world
constraint system is a finite, non-empty set of world constraints.

A Kripke structure I< = (W, r, I<I) satisfies a world constraint system W iff for
each label w in W there is a world w K E W such that (i) I<, w K 1= F for each world
constraint F II win Wand (ii) (wK,v K ) E Ii for each world constraint w ~i v in W.
A world constraint system W is (KD45- ) satisfiable iff there exists a (KD45) Kripke
structure which satisfies W.
For testing KD45-satisfiability of a set FI, ... ,Fn of ACCs-formulas we firstly trans-
late them into a world constraint system. The world constraint system W is induced by
F1 , . , Fn iff W = {F1Ilwo, ... , Fn IIwo}, where Wo is a new constant (which represents
the real world). 0 bviously, F1 , . , Fn are KD45-satisfiable iff W is KD45-satisfiable.
KD45-satisfiability of a world constraint system W is tested by the frame algorithm
which has a world constraint system as input that is induced by a finite number of
ACCs-formulas and which successively adds new world constraints to W by applying
the four propagation rules in Figure 1. Thus, the result of the frame algorithm with
input W is a (modified) world constraint system W'. We will call each world constraint
system that can be obtained from W by a finite number of propagation rule applications
a derived system in the following .
The intuitive idea behind these propagation rules is as follows: Firstly, for W -+0
W', if there is a world constraint OiF II w in W we add a world v such that (i) v is
accessible from w by agent i and (ii) F II v holds. Furthermore, whenever OiFj II w
is in W we add OiFj II v because of property (P;), and whenever DiGk II w is in W
we add both DiGk II v and G k II v because of property (P 2 ) and the definition of D i .
This rule is similar to the unsigned prefixed KD45 tableaux rules in [14]. Secondly, for
W -+0 W', if DiG l II w, . .. , DiG m II ware in W but there is no world u accessible from
w by agent i, we have to introduce a new world v-accessible from w by agent i-where
G 1 II v, ... ,Gm II v and DiG l II v, ... ,DiG m II v holds. This is due to the properties
12
W --+0 {w ~i V, Fllv, oiFlllv, ... , OiFn Ilv, Glllv, DiGlllv, ... , G m Ilv, DiG m Ilv }UW
if OiF II w, OiFl II w, ... ,OiFn II ware the world constraints with leading
modality 0i in W, DiGl II w, ... , DiGm II ware the world constraints
with leading modality Di in W, there is no label u in W such that the
world constraints F II u, OiFl II u, . .. , OiFn II u, G l II u, ... , Gm II u,
DiGl II u, ... ,DiGm II u are exactly the labeled ACCB-formulas with
label u in W, and v is a new label.
W --+0 0 {w ~i u} U W
if OiF II w, OiFl11 w, .. . , OiFn II ware the world constraints with leading
modality Oi in W, DiG I II w, ... , DiGm II ware the world constraints
with leading modality oi in W, there is a label u in W such that the
world constraints F II u, 0iFl II u, ... ,OiFn II u, G I II u, ... ,Gm II u,
oiG l II u, . . . ,DiGm II u are exactly the labeled ACCB-formulas with
label u in W, and w ~i u is not in W
W --+0 {w ~i v, G l II v, DiG l II v, ... , Gm II v, OiGm II v} U W

if no world constraint of the form OiFllw is in W, DiGlllw, ... ,DiGm Ilw
are the world constraints with leading modality Di in W, there is
no label u in W such that the world constraints G 1 II u, ... ,Gm II u,
DiGl II u, .. . , DiGm II u are are exactly the labeled A..cC 6 -formulas with
label u in W, and v is a new label.
W --+0 0 {w ~i u} U W
if no world constraint ofthe form OiFllw is in W, DiGlllw, ... , DiGm Ilw
are the world constraints with leading modality oi in W, there is a
label u in W such that G l II u, ... , Gm II u, DiGl11 u, ... , DiGm II u are
are exactly the labeled ACCB-formulas with label u in W, and w ~i u
is not in W.
Figure 1: Propagation rules of the ACCB frame algorithm.
(P~) and (P 2 ) of KD45 Kripke structures. Finally, the rules --+0 and --+0 are used to
0 0
guarantee termination of applying propagation rules to a derived system.

Now we will show that the ACC B frame algorithm has the following two important
properties. Firstly, it terminates for every world constraint system W as input which
is induced by a finite set FI,"" Fn of ACCB-formulas. Secondly, if W' is the result
of the frame algorithm with input W, then F I , ... , Fn are KD45-satisfiable iff for each
label w in W'the set of ACC-formulas with label w in W' is satisfiable.
13
Termination of the frame algorithm is stated by the next proposition. Its proof
will employ techniques which have been developed for proving termination of term
rewriting systems (see [11]).
Proposition 3.3 If W is a world constraint system which is induced by a finite set of

Ales-formulas, there is no infinite chain of applications of propagation rules to W.
In order to prove this proposition we will map derived systems to multisets which
are equipped with a well-founded strict partial ordering~. Multisets are like sets,
but multiple occurrences of elements are allowed, e.g., {2, 3, 3, 4} is a multiset which is
different from the set {2,3,4}.
A given ordering> on elements in a set S can be extended to an ordering on~
finite multisets over S as follows. If M and M' are finite multisets over S then M ~ M'
iff M' is obtained from M by replacing one or more elements in M by a finite number
of elements in S, each of which is smaller than one of the replaced elements (w.r.t. .
For example, {2,3,3,4} is larger than the multisets {2,3,1,2,3} and {4}. Dershowitz
and Manna [11] showed that the ordering ~ on finite multi sets over S is well-founded
if the original ordering on S is so.
We will use a mapping \lI which maps derived systems to multisets whose elements
are pairs of non-negative integers. These pairs are ordered lexicographically from left
to right , i.e., (CI, C2) is larger than (c~, c;) iff (i) CI > S or (ii) CI = c~ and C2 > c;.
Since the orderings on both components are well-founded, the lexicographical ordering
on these pairs is also well-founded. Finite multisets of these pairs are now compared
w.r.t. the multiset ordering which is induced by this lexicographical ordering. This is
the well-founded ordering ~ mentioned above.
In order to simplify notation of the mapping \lI we introduce the following notions.
If F II wand G II ware labeled Ales-axioms with the same label w, we say that
G II w is a modal subformula of F II w iff there is a (possibly empty) sequence o of
modalities such that o"G and F are identical. For example, OjOiOjF II w, 0iOjF II w,
OjF II w, and F II ware all modal subformulas of OjOiOjF II w. If F II w is a labeled
Ales-formula we denote the set {G I Gil w is a modal subformula of F II w} by MSub
(F II w). For a derived system W, MSubW(w) denotes the set UFllwEW MSub (F II w).
Now we can define the mapping \lI as follows.
Definition 3.4 Let Wo be a world constraint system which is induced by a finite num-
ber of Ales-formulas, let W be a world constraint system which is derived from Wo
by applications of the propagation rules of the Ales frame algorithm, and let w be a
label in W. We define
1. S is the power set of {MSubwO{wo) I Wo is the (only) label in W o}, and Ns is the
number of elements in S.
14
2. Nt' is the number of labels in W which are different from woo
3. 1f1r'(w) is the number of modalities in the set {F II w I F II w E W}.
4. 1f1'f (w) is the cardinality of {w ~j v I w ~j v E W for some agent j and label v} .
5. The mapping 1f1w (w) is now defined by the pair
of integers, and W(W) is the multiset {t/;w (w) Iw is a label in W}.
Given a derived system Wand a label win W we firstly show that both components of
t/;W(w) are non-negative integers. It is easy to verify that each of the four propagation
rules only introduces labeled A.cCB-formulas HI II v, .. . , Hn II v such that each Hi is
an A.cCB-formula in MSubwO(wo). In other words, the set {HI,"" Hn} is an element
in S. Because of the definition of the propagation rules a new label v-together with
some labeled A.cCB-formulas HI II v, ... , Hn II v-is only introduced if there is no label
u in W such that the world constraints HI II u, ... ,Hn II u are exactly the labeled
A.cCB-formulas with label u in W. That means, for each element in S at most one
new label can be introduced, and hence Ns - Nt' ~ 0 for each derived system W.
On the other hand, since each propagation rule can be applied to a modality in a
labeled A.cCB-axiom at most once (because of the disjoint preconditions of the four
propagation rules) and adds exactly one world constraint of the form w ~i v, we can
conclude that 1f1r' (w) - t/;'f (w) ~ 0 for each label w in W.
We will now show that each chain Wo -+1 WI -+2 ... -+n Wn of propagation
rule applications to derived systems corresponds to the decreasing chain w(Wo) ~
\l!(Wl) ~ ... ~ \l!(Wn)' Thus, proposition 3.3 is an immediate consequence of the
next lemma.
Lemma 3.5 If W' is obtained from the derived system W by an application of a prop-
agation rule, then W(W) ~ \l!(W').
Proof: We have to show this lemma for each of the propagation rules.
(1) Assume that W' is obtained from W by applying the -+0 or the -+0 rule to
one or more world constraints labeled with label w. Such a rule application adds a
world constraint w ~j v to W, where j is an agent and v is a new label. Hence, the
number of labels in W' is greater than the number of labels in Wand for each label
u in W'the first component of 1f1w'(u) is less than the first component of 1f1W(w), i.e.,
\l!(W) ~ W(W').
(2) Now assume that W' is obtained from W by an application of the -+0 or 0
the -+0 rule. Such a propagation rule application does not introduce a new label
0
15
to W. Thus, the first components of WW(u) and WW'(u) are identical for each label
u. Let us now consider how the values ~r' (w) and ~'f (w) are changed. Obviously,
~r' (w) = ~r" (w) and ~'f (w) = ~'f' (w) + 1. From this it follows immediately that
WW(w) ~ WW'(w). 0
Thus, the application of the frame algorithm to a world constraint system W in-
duced by a finite set of ACB-formulas Fb ... , Fn terminates and results a world con-
straint system, say W'. In order to test KD45-satisfiability of W', for each label w in
W' we determine the set of all those ACB-formulas which are labeled by wand which
do not contain any modality. That means, such a set contains only AC-formulas
and is therefore called the AC test set of label w. More formally, if W' is a world
constraint system, the AC test set A( w) of label w in W' is given by the set
{F I F II w E W' and F does not contain any modality}.
We will show in the following that a finite set Fb"" Fn of ACB-formulas is KD45-
satisfiable iff the AC test set A( w) of each label in W' is satisfiable. Thereby, W' is
the result of the frame algorithm with input {Fl II Wo, ... , Fn II wo}. One direction is
given by the next lemma.
Lemma 3.6 Let W be a world constraint system which is induced by the finite set
{F1 , .. , Fn} of ACB-formulas, and let W' be the result of the frame algorithm with
input W. If K = (W,r,Kl) is a KD45 Kripke str1f,cture which satisfies W, then for
each label w in W' there is a world wK E W such that K, wK F F for each labeled
ACB-formula F II w in W'.
Proof Since W' is the result of the frame algorithm with input W there is a chain W =
Wo -tl WI -t2 ... -tk W k = W' with -tiE {-to,-too,-to,-toJ for i E {I, .. . ,k}.
We will show that K satisfies each labeled ACB-formula in W' by induction over the.
number of rule applications in this chain. By assumption, K = (W, r, K 1 ) satisfies
W o = {F1 II wo, ... ,Fn II wo}, i.e., there is a world w{f in W such that K,w{f F
Fl"'" K, w{f F Fn
We thus can assume that, after j rule applications, for each label w in Wj there
is a world w K in W such that K, w K F F for each labeled ACB-formula F II w in
Wj . If Wj -tj Wj+! there are four possibilities. Firstly, suppose Wj -to Wj+! by
applying the -to rule to 0 i F II win Wj. In this case, there are labeled ACB-formulas
0 i F II W, 0 i Fl II w, ... , 0 i Fn II w, DiGl II w, ... , DiGm II win Wj, and Wj +1 is given by
where v is a new label. By induction hypothesis there is a world w K in W such that

(i) K,w K F 0 i F and (ii) K,w K F DiGj for j = 1, ... ,m. Because of (i) there
is a world v K in W (not necessarily different from w K ) such that (w K , v K ) E Ii and
16
K, V K 1= F. Furthermore, because of (ii) and property (P2) of KD45 Kripke structures,
both K, w K 1= DiGj and K, w K 1= DiDiGj holds for j = 1, ... , m. And thus, since
(w K , v K ) E Ii, especially K, v K 1= G j and K, v K 1= DiGj holds for j = 1, ... , m.
Finally, K, w K 1= OiFj for j = 1, ... ,n by induction hypotheses, i.e., K, w K 1= DiOiFj
because of property (P3') of KD45 Kripke structures. Hence, also K, v K 1= OiFj holds
for j = 1, ... , n. Summing up , for each labeled ACB-formula F II v in Wj we have
K, v K 1= F .
Secondly, if Wj ~D W j +1 there are world constraints DiGl11 w, ... , DiGm II win Wj
and W j +! = Wj U {w txl i v, Glll v,DiGlll v, ... ,Gm II v,DiG m II v} where v is a new
label. By induction hypothesis , K, w K 1= DiG j for j = 1, ... , m and for some world
w K E W. Because of K , w K 1= DiG l and property (PI') of KD45 Kripke structures
especially K, w K 1= OiG 1 holds. Hence, there is a world v K (not necessarily different
from wK ) such that (w K , v K ) E Ii and K, v K 1= G 1 Furthermore, because of property
(P2) of KD45 Kripke structures, both K, w K 1= DiGj and K, w K 1= DiDiG j holds for
j = 1, ... ,m. And thus, since (w K , v K ) E Ii, especially K, v K 1= G j and K, v K 1= DiG j
holds for j = 1, ... , m. Summing up, we have K, v K 1= DiG j and K , v K 1= G j for
j = 1, .. . , m .
Finally, if Wj ~oo W j +1 or Wj ~Do W j+! there is nothing to show since these rules
do not change the set of labeled ACB-formulas in Wj at all. D
The next lemma states that a world constraint system W', which is a result of
the frame algorithm, is KD45-satisfiable if the A.ce test set of each label in W' is
satisfiable.
Lemma 3.7 Let W be a world constraint system which is induced by a finite set of
A.cCB-formulas, and let W' be the result of the frame algorithm with input W. If the
A.cC test set A( w) of each label w in W' is satisfiable, then W is KD45-satisfiable.
Proof: Let K be the Kripke structure (W, f, K 1 ) where
W is given by the set of all labels in W'.
f consists of an accessibility relation Ii for each agent i. Thereby, Ii is given by

the Euclidean and transitive closure of the set {( w, v) I W txl i v in W'} U {( w, w) I
w is a label in W' and for no label v in W' there is w txl i vin W'}. It is easy to
verify that each accessibility relation Ii is serial.
KI is given such that K , w 1= F for each labeled A.cCB-formula F II w in W'

where F does not contain any modality. Such a K-interpretation K I exists,
since we assumed the A.cC test set of each label in W' to be satisfiable. Given
interpretations II, ... ,In which satisfy the A.cC test sets of each label in W'
respectively, the construction of K I is straightforward.
17
Obviously, each /i is Euclidean, transitive, and serial and hence I< is a KD45 Kripke
structure.
We will now prove that I< satisfies each world constraint c in W'. If c is of the form
w IXl j v there is nothing to show because of the defini tion of I<. The fact I< F F II w for
each labeled ALes-formula F II w in W' can be shown by induction over the number
of modalities in F. If F does not contain any modality, then I<, w F F because of the
construction of I<. Thus we can assume that I<, w 1= F for each labeled ALes-formula
F II w in W' such that F contains n modalities.
If F contains n + 1 modalities, there are two possibilities: the leading modality is
either 0i or <\. Firstly, suppose W' contains a world constraint 0iF II w, where F has
n modalities. We then have to show that I<, w 1= 0iF, i.e., that I<, v 1= F for each v
such that (w,v) E Ii. Since OiFllw is in W', during the frame algorithm a propagation
rule has been applied to the world constraints with label w, such that w lXl i u is in W'
for some label u. It is easy to verify that w lXl i v E W' for some label v (not necessarily
different from w) if OiF II w or 0 i F II w is in W' for some ALes-formula F. Thus,
because each /i is transitive and Euclidean, there are two possibilities for (w, v) E /j:
2. there are two paths starting with some label z, namely
For case 1., assume that W' contains a world constraint OjF II w. Because of the
definition of the propagation rules it holds that, whenever DiF II w is in W' and w lXl i v
(or w IXl j w) is added to some derived system W j , then both F II v and DjF II v are in
Wj. Analogously, whenever a world. constraint v IXl j u is added to a system Wjt (with
j' 2 j + 1), the derived system Wjt contains DiF II u and F II u, and so on. Hence, since
none of the propagation rules deletes a world constraint, we can conclude that F II v is
in W' if there is a path w = Wi l IXl j Wi2' Wj2 IXl j Wi 3 , ,Wjk_l IXl j Wjk = v in W'. And, by
induction hypothesis, we know that I<, v F F because F contains only n modalities.
For the second case, assume the two above given paths starting with label z are in
W'. We firstly show that
(*) if DiF II w is in W', then DjF II z is also in W'.
It is sufficient to show that DiF II u is in Wj whenever DiF II u' is in Wj + b where Wj - 0

Wj +1 , Wj - 0 Wj +b Wj - 0 Wj +b or Wj - 0 Wj +1 , and u IXl j u' is added to Wj by
0 0
this rule application. This holds because of the definitions of the four propagation rules
and since ALes-formulas are in KD45 normal form, such that (*) follows immediately.
18
1. Let W be the world constraint system which is induced by Fl, ... ,Fn.
2. Let W' be the result of the frame algorithm with input W.
3. For each label w in W' do: If the AC test set of w is not satisfiable, then
STOP and return "KD45-unsatisfiable".
4. Return "KD45-satisfiable".
Figure 2: The KD45-satisfiability algorithm.
That means, if DiF II w is in W' we know that DiF II z is in W'. As argued above, in
this case W' contains the labeled ACs - formula F II v because of the path z [Xli [Xli v.
Again, we know I<, v F F because of the induction hypothesis.
Secondly suppose W' contains 0 i F II w. We then have to show that I<, v F F for
some world v such that (w, v) E Ii. This is obvious , since either (i) the ---+0 rule has
been applied to 0 i F II wand added both world constraints w [Xli v and F II v for some
label v, or (ii) the ---+0 rule has been applied to 0 i F II wand added w [Xli u to W' such
0
that F II u E W'. Summing up, I< satisfies each world constraint in W' and hence in
W ~ W'. Thus, W is KD45-satisfiable. 0
The next theorem summarizes the previous results.
Theorem 3.8 Let F1, .. . , Fn be a finite set of AC 13 -formulas, and let W be the world
constraint system which is induced by F 1, . .. , Fn. If W' is the result of the frame
algorithm with input W , then the set F 1 , .. ,Fn is KD45-satisfiable iff the AC test
set A( w) of each label w in W' is satisfiable.
Proof; The ACs-formulas F 1 , . , Fn are KD45-satisfiable iff W is KD45-satisfiable.

Firstly, suppose I< is a KD45 Kripke structure which satisfies W . Then, because of
Lemma 3.6, for each label w in W' there is a world w K E W such that I<, w K 1= F for
each ACs-formula F II w in W'. Thus, especially the AC test set of each label w in
VV ' is satisfied by I<. Conversely, suppose that the AC test set A( w) of each label w
in W' is satisfiable. Then W is KD45-satisfiable because of Lemma 3.7. 0
Summing up, we obtain the algorithm for testing KD45-satisfiability of AC s -

formulas F1, ... ,Fn which is given in Figure 2. An algorithm for testing satisfiability
of AC test sets has been given in [21J.
Unfortunately, the number of labels in the constructed world constraint system is in
the worst case exponential in the number of input AC13 -formulas: Let 0 1 F 1 , . . . , 01Fn
be ACs-formulas where each Pi is an AC-formula. The induced world constraint
19
system Wo is then given by
Applying the - 0 rule successively to 01F1 II Wo,, OlFn II Wo results the world con-
straint systems
WI Wo U {wo ~l WI, Fl II w}, 01F211 WI,, OlFn II wd

W2 WI U {wo ~l w2, 01F1 II W2, F211 W2, 01F3 11 W2,, OlFn II W2}
Each of the world constraint systems WI, ... , Wn contains n -1 labeled A.cCB - formulas
with a leading modality 0 1 , i.e., to each of these derived systems the - 0 rule can be
applied n -1 times. 4 To each of the thereby obtained derived systems the - 0 rule can
be applied n - 2 times and so on. summing up, an exponential number of labels-and
hence of labeled A.cCB-formulas-is generated.
In order to test KD45-satisfiability of a set of A.cCB-formulas we are mainly inter-
ested in the constructed A.cC test sets which have to be tested on satisfiability. The
following example shows that the number of different A.cC test sets in a derived system
W may essentially be smaller than the number of labels in W.
Example 3.9 Let the A.cCB-formulas OlFI, 01F2, OlF3 be given where F1, F2, F3 are
A.cC-formulas. Applying the frame algorithm to
results a world constraint system W' which has 13 different labels. H oweverJ only three
different A.cC test sets {namely {FdJ {F2L and {F3}) are constructed.
An optimized version of the KD45-satisfiability algorithm-which does not generate a

worst case exponential overhead of labeled A.cCB-formulas-is presented in the follow-
ing section.
4 Optimization and Computing ALes-Inferences
In this section we consider optimizations of the KD45-satisfiability algorithm as well

as the problem of computing logical inferences from given A.cCB-formulas. In 4.1 we
4For sake of simplicity we disregard the - 0 rule here. It is easy to verify that applications of this
rule do not influence the exponential behaviour of the frame algorithm with input OlF1 ,.., OlFn .
20
will present an algorithm which-based upon the results of the previous section-
determines ACC test sets without computing an exponential number of labeled ACCB -
formulas. These sets then have to be tested on satisfiability, i.e., we have to decide
whether or not a set of terminological and assertional axioms is satisfiable. Termi-
nological axioms are defined to be of the form C = D or C =1= D where C, Dare
(complex) concepts. However, in most of the existing terminological representation
systems terminological axioms are not allowed in this general form but they have to
satisfy additional conditions. In 4.2 we investigate in which cases only such restricted
terminological axioms have to be considered in order to test satisfiability of ACC test
sets. Finally, in 4.3 we will show how to decide whether or not a given ACCB-formula
is logically entailed by a set of AlCB-formulas.
4.1 Computing A.ce test sets
The KD45-satisfiability algorithm presented in the previous section works in two phas-
es: Firstly, a set of AlC test sets is computed and then each of these sets is tested
on satisfiability. The thereby used propagation rules of the frame algorithm have the
advantage to "reflect" the properties of KD45 Kripke structures, and hence soundness
and completeness of the KD45-satisfiability algorithm could be proved in a very natural
way. On the other hand, the frame algorithm has the disadvantage to possibly con-
struct a large number of labeled A.L:es-formulas in order to determine a small number
of ACC test sets (d. example 3.9).
In the following we will have a closer look at the properties of the frame algorithm.
Building upon these properties we will then develop an algorithm which computes A.L:C
test sets immediately from the ACCB-formulas to be tested on KD45-satisfiability, i.e. ,
we do no longer have to compute a (large) number of labels from which ACC test sets
then are extracted.
The main idea of this new approach is as follows: Suppose oi F1 , .. , o~Fn are ACC B -
formulas which are to be tested on KD45-satisfiability, where each 07 is a (possibly
empty) sequence of modalities and each Fi is an ACC-formula. By looking at the
structure of the sequences oi, ... ,o~ we will then decide syntactically which elements
in the power set of {Ji1, . .. , Fn} will be computed as an ACC tes t set by the frame
algorithm with input 0iFl II Wo, ... , o~Fn II woo The following example shows that this
task is not obvious.
Example 4.1 Let A be a primitive concept and let C, D be concepts.
1. Let S be the set {Ol(A = C),Ol(A = D)}. Applying the frame algorithm to
the induced world constraint system W = {Ol(A = C) II wo,Ol(A = D) II wo}
21
results the derived system W', given by the world constraints
OI(A = C) II Wo OI(A = D) II Wo
Wo t><II WI A = C II WI OI(A = D) II WI
Wo t><II W2 OI(A = C) II W2 A = D II W2
WI t><II W3 A = D II W3
W2 t><II W4 A = C II W4
The ACC test set A(wo) of Wo is empty, while A(WI) = A(W4) = {A = C} and
A(W2) = A(W3) = {A = D}.
2. On the other hand, starting with the set S = {DI(A = C), DI(A = D)} leads to
only one non-empty ACC test set, namely {A = C, A = D}.
The leading sequences of modalities in the input ACCB-formulas obviously influence

the determined ACC test sets.
Let us firstly introduce the following notions. If W is a derived system, a subset

consisting of world constraints WI t><Iil W2, ... , Wn-I t><Iin_1 Wn is called a path if the Wj
are labels in Wand each i j is an agent. The norm of this path is given by the sequence
iI' ... in' which arises from i l ... i n - I be replacing each subsequence i j ... i j in i l ... i n - l
by i j . For example, the norm of the path WI t><Iil W2, W2 t><Iil W3, W3 t><Ii2 W4, W4 t><Ii2 W5 is
given by i l i 2 .
In the following we will present correlations between labeled ACCB-formulas which
are taken as input of the frame algorithm and the labeled ACCs-formulas which are
generated during the frame algorithm with this input. In order to simplify notation
we thereby assume the input is of the form O~FI II Wo, ... , o~Fn II Wo where each 07
is a (possibly empty) sequence of modalities and each two ACC-formulas Fi and Fj
are syntactically different (e.g., by marking them with numbers). Since the presented
correlations only depend on the modalities in the input ACCs-formulas this can be
done without loss of generality.
Firstly, if W' is the result of an application of the frame algorithm, the next lemma
states that there is a correlation between paths in W' and the sets of ACCB-formulas
in W'.
Lemma 4.2 Let W be a world constraint system which is induced by a finite set of
ACCs-formulas, let W' be the result of the frame algorithm with input W, and let P
be a path Wo t><Iil WI, ... , Wn-l t><I in Wn in W' with norm (P) = i l ... in. Then the
set of labeled ACCB-formulas in W' with label Wn is a subset of {F II Wn , Oi n F II Wn I
Oil. 0i n F II Wo E W'} where E {D,O}.
22
W'. If OiFl "wo, . .. ,0iFn " 100, DiG! II Wo, . . , oiG m II Wo are
Proof: Let Wo t><I i u be in
the labeled ACB-formulas in W' with label Wo and with leading modality oi or OJ,
the labeled ACB-formulas in W' with label u are a subset of 5, given by
or, alternatively, by {F II u, OiF II u I OiF II Wo E W'}. This follows immediately from

the definition of the four propagation rules. Since we assumed ACB-formulas to be in
KD45 normal form, none of the world constraints Fl II U, ... , Fn II u, G 1 II U, , .. , G m II u
has a leading modality oi or Oi. Hence, if for some label u' there is u t><I i Wk we know that the set of labeled AC B-
formulas with label Wk in W' is a subset of 5 = {F II Wk, Oi F II Wk I Oi F II Wo E W'}.
Let us now consider a path P with norm (P) = ij. If there is a path Wo t><I j Wjl' then the set of labeled ACB-formulas with label Wk
in W' is a subset of 5 = {F II Wk, OiF II Wk I Oi F II Wo E W'}. Hence, the set of labeled
AI.CB-formulas with label Wjl in W' is a subset of {F II Wjl' ojF II Wjl I ojF II Wk E 5}.
Analogously to the argumentation above we obtain the following: If there is a path
Wo t><I j WI in W' it follows that the set of
labeled ACB-formulas with label WI in W' is a subset of {Fllwl, OjFllwl I OjFllwl; E 5}.
Since 5 consists only of elements OiF II Wk or F II Wk such that OiF II Wo is in W' it
follows that ojF II Wk E S .at most if 0i OJ F II Wo E W'. That means, the set of labeled
ACB-formulas with label WI in lV' is a subset of {YII WI, ojF II w/ I 0i OJ F II Wo E W'}.
With this argumentation the lemma follows immediately by induction. o
In an AC test set only AC-formulas, i.e. A.cCB-formulas without modalities, do
occur. If W is a world constraint system and W' is the result of applying the frame
algorithm to W, the next lemma provides a correlation between labeled ACB-formulas
in W' without modalities and labeled ACB-formulas in W. Thereby, a sequence of
modalities is called sequence for short. If 5 is the sequence 01 . . . On, with E {O, o},
then 5[j] denotes OJ and indexes (5) denotes 1, ... ,n. Furthermore, an ACB-formula
F occurs in a derived system iff there is a (possibly empty) sequence 0* and a label W
in W such that 0* F II W is in W.
Lemma 4.3 Let W be a world constraint system which is induced by a finite number
of ACB-formulas, let W' be the result of the frame algorithm with input W, and let F
and G be Ae-formulas occurring in W. If there is a label Wn in W' such that F II Wn
and G II Wn are both in lV' , then there are two sequences 5 and 5' such that (i) 5F II Wo
and 5'G II Wo are in Wand (ii) indexes (5) = indexes (5').
Proof: Let Wo t><Iil WI, ... , W n -1 t><IiJ Wj has been added into W' by an application of the -to or the -to rule. It is
23
sufficient to investigate this path since the -too and the -too rule do not add any new
labeled ALCs-formulas at all, and it is easy to verify that there is exactly one such
path.
If Fllw n and Gllwn are both in W', there are labeled ALCs-formulas oil" .0inFllwo
and oil'" 0i n G II Wo in W where each 0 is either 0 or
and each i j is an agent. This
follows immediately by Lemma 4.2 and the fact that H II Wo E W' iff H II Wo E W for
each ALCs-formula H since no propagation rule introduces a labeled ALCs-formula
with label Wo into a derived system. Thus, there are sequences 5 and 5' such that
indexes (5) = indexes (5') and both 5F II Wo and 5'G II Wo are in W.
If W' is the result of applying the frame algorithm to a world constraint system
W, the following proposition provides syntactical conditions that W satisfies whenever
there is an ALC test set A( w) for some label W in W which contains a non-empty set
of ALC-axioms.
Proposition 4.4 Let W be a world constraint system which is induced by a finite

number of ALCa-formulas, let W' be the result of the frame algorithm with input W,
and let F and G be ALC -formulas occurring in W. If there is a label Wn in W' such
that F II Wn and G II Wn are both in W', then there are sequences 5 and 5' such that (i)
indexes (5) = indexes (5'), (ii) 5F II Wo and 5'G II Wo are in W, and (iii) there is no
position j in 5 (respectively in 5') such that 5[j] = 5'[j] = O ij for some agent i j
Proof Because of Lemma 4.3 we know that there are sequences 5 and 5' with index-
es (5) = indexes (5') and both 5F II Wo and 5'G II Wo are in W. Suppose j is the first
position such that 5[j] = 5'[j] = O i].
Starting with label Wo we firstly show that for each path Wo !><lil WI, . . , Wn-I !><lin Wn
in W' there is a label Wj in Wo, ... ,Wn such that Oij oi F II Wj and O ij O 2G II Wj are in
W', where o~ and 0; are (possibly empty) sequences of modalities and indexes (o~) =
indexes (0;). Therefore we use the fact that for each position k < j at least one of the
modalities in {5[k], 5'[k]} is not a labeled O-modality.
If O ij oi F II Wo and O ij O 2Gil Wo are both in W' there is nothing to show. Otherwise,
we have to distinguish two cases:
(i) O i 0* O ij F II Wo and Di 0* Oij O 2G II Wo are in W'

O~ or
(ii) Di 0* Oij oi F II Wo and 0i 0* Oil O 2G II Wo are in W',
where i is an agent and 0* is a (possibly empty) sequence of modalities. s Without loss

of generality we do not consider the case where Di 0* Oij oi F II Wo and Oi 0* Oij O 2G II Wo
5In order to simplify notation we will use o in both the sequences of F and of C. Formally we
had to distinguish these occurrences since everything we know about them is that they do not differ
in their indexes. However, here it should always be clear by context which occurrence is meant .
24
are in WI, since it is symmetrical to case (i). Note, that applying propagation rules to
labeled .A.cCs-formulas with label Wo whose leading modalities are not indexed with
i-obtaining, say, label v-cannot result the first element in a path leading to a label
w such that F II wand G II ware both in WI. This is due to the fact that in this case
neither F nor G occurs in labeled .A.cCs-formulas labeled with v.
For case (i), after an application of the ---+0 or the ---+0 0 rule to OJ 0* Oij o~ F /I Wo,
for some label v the world constraints
o*Ojj o~ F /I v, o*Oj) 0; G /I v, OJ 0* Oij 0; G /I v (a)
are exactly the labeled .A.cCs-formulas in the currently derived world constraint system
which are labeled with v and contain F or G. Analogously, after an application of the
---+0 or the ---+0 rule to OjH II wo-where H is different from O*Oi) o~ F-for some label
0
v the world constraints
are exactly the labeled A.cCs-formulas in the currently derived world constraint system
which are labeled with v and contain F or G.
For case (ii), after an application of the ---+0 or the ---+0 rule to labeled .A.cCs - 0
formulas with label Wo whose leading modalities are indexed by i, for some label v the
world constraints
OJ 0* OJ) o~ F /I v, O*Oi) o~ F /I v, OJ 0* Ojj 0; G /I v, O*Oij 0; G /I v (J)
are exactly the labeled A.cCs-formulas in the currently derived world constraint system
which are labeled with v and contain F or G. And, finally, after an application of the
---+0 or the ---+0 rule to some labeled A.cCs-formula OjH /I Wo we obtain the same
0
labeled A.cCs-formulas as in case (J) above.

Since we assumed A.cCs-formulas to be in KD45 normal form, the leading modality
in 0* is not indexed with i. Hence, for case (a), propagation rule applications to
OJ 0* Oi) 0; G II v cannot result the first element in a path leading to a label w such that
F /I wand Gil ware both in W I. Analogously, for case (13), applying propagation rules
to O*Oi) O2 G /I v cannot lead to a label w such that F /I wand G /I ware both in WI.
For the same reason in case (J) either rule applications to labeled .A.cCs-formulas with
leading modality i and label v, or to labeled .A.cCs-formulas whose leading modality is
the first modality in 0* and label v, may lead to a label w such that F II wand Gil w
are both in WI.
Summing up, if v is the first element in a path Wo [Xljl' . . . ,[Xli n Wn , then either
(1) o*Ojj o~ F II v and o*Ojj 0; G II v are in WI or

(2) Oi 0* OJ) o~ F /I v and OJ 0* Oij 0; G II v are in WI or
(3) OJ 0* OJ) o~ F II v and OJ 0* Oij 0; G II v are in WI
25
Hence, to v the same argumentation is applicable as to label Wo above.
Let now Wj be a label in wo, ... , Wn such that 0 ij oi' F " Wj and 0 ij 0; G II Wj are in
W' . Without loss of generality during the frame algorithm the - 0 (respectively the
-0.) rule has been applied to 0 ij oi' F II Wj before it has been applied to 0 ij o~ G II Wj.
If oi' (and hence O 2) is the empty sequence, then F II v and 0 ij G II v are in W', where
Wj ~ij v has been introduced by this rule application. But F II Wn cannot be in W'
because of the remaining application of the - 0 rule to 0 ij o G II Wj. On the other
hand, if oi' (and hence o~) is not empty, its leading modality is different from i j (since
A.cCB-formulas are in KD45 normal form) and thus there cannot be a label Wn in W'
such that F II Wn and G II Wn are both in W' because of the remaining application of
the - 0 or the -00 rule to 0 ij o G II Wj.
In both cases this contradicts the assumption that there is a label Wn in W' such
that F II Wn and Gil Wn are both in W', i.e., there cannot exist a position j such that
5[j] = 5/[j] = 0 ij for some agent i j 0
Finally, we present a proposition which states that also the opposite direction of
proposition 4.4 holds.
Proposition 4.5 Let W be a world constraint system which is induced by a finite set
of A.cCB-formulas, let W' be the result of the frame algorithm with input W, and let F
and G be A.cC-formulas which occur in W . If 5F II Wo and 5 /G II Wo are in W , where
5, 5' are sequences such that indexes (5) = indexes (5') and there is no position j in 5
(respectively in 5') such that 5[j] = 5 / [j] = 0 ij for some agent i j , then there is a label
Wn in W' such that F II Wn and G II Wn are both in W'.
Proof: Let 5. and 5: arise from 5 and 5' by deleting the first modality, respectively. If
5F II Wo and 5/G II Wo are in W, there are three cases:
1. 5[1] = 5 /[1] = Oil and the - 0 or the - 0 0 rule is applied to A.cCB-for~ulas of
the form O ilH II woo In this case 5F II WI, 5 /G II WI, 5.F II WI, and 5:G II WI are
all in W' if Wo ~il WI is introduced by this rule application.
2. Without loss of generality 5[1] = 0 ill 5/[1] = Dill and the - 0 or the - 0 0 rule
is applied to some A.cCB-formula 0 il H II Wo, where H is different from F. In
this case 5F II WI and 5 /G II WI are in W' if Wo ~il WI is introduced by this rule
application.
3. Without loss of generality 5[1] = 0 i l , 5 /[1] = Oil, and the - 0 or the - 0 0 rule is
applied to 5Fllwo. In this case 5.Fllwi and 5:Gllwi are both in W' ifwo ~il WI
is introduced by this rule application. It does not influence the argumentation
that 5 /G II WI is also in W'.
26
For label WI the argumentation is the same such that the proposition follows immedi-
ately by induction. 0
Propositions 4.4 and 4.5 together give us an answer to the question which ALe test
sets are generated by an application of the frame algorithm. We summarize this result
in the following theorem.
Theorem 4.6 f Let W be a world constraint system which is induced by a finite set
of ALes-formulas and let W' be the result of the frame algorithm with input W. Then
there is a label W in W' such that the ALe test set A( w) contains the ALe -form ulas
F l , ... ,Fn iff there are sequences 51, ... ,5 n in W such that
indexes(5 1 ) = ... = indexes(5 n ), and
there is no position j such that for two sequences 5' and 5" in {51, ... ,5 n } holds
5'U] = S"[j] = Oil for some agent i j
Proof; If the ALe test set A( w) contains the ALe-axioms FI, . .. ,Fn , none of the F j
does contain modalities. Thus, because of Proposition 4.4 there are sequences 51, ... ,5 n
such that SIFl II Wo, . .. ,SnFn II Wo are in Wand indexes(St) = ... = indexes(Sn). Now
suppose, that 5 1 F'11 Wo and 5"F"" Wo with F' and F" in {Fl , . . . ,Fn} are in W such
that 5/[j] = 5"[j] = Oil for some agent i j . Again because of Proposition 4.4, we can
conclude that either F' " w or F" " w is not in W'. This contradicts that A( w) contains
F l , . . . , Fn. The other direction follows immediately from Proposition 4.5. 0
An optimized version of the KD45-satisfiability algorithm is given in Figure 3. For

the input of this algorithm remember that each ALes-formula is of the form 5F where
S is a (possibly empty) sequence of modalities and F is an ALe-formula. In step 2.
of the algorithm for each sequence 5i in {51, ... , 5n } the set Si of all ALe-formulas
in {Fl , . . . , Fn} is computed such that the conditions of Theorem 4.6 are satisfied.
Multiple generations of the same ALe test sets are avoided by testing whether or not
the ALe-formula Fi is already in an ALe test set Sk.
Summing up, soundness, completeness, and termination of this algorithm follow
immediately from the results in Section 3 and in this subsection. It is easy to verify
that step 2. of this optimized KD45-satisfiability algorithm works in polynomial time
in the length of the input ALes-formulas and generates a polynomial number of ALe
test sets. Thus, in contrast to the KD45-satisfiability algorithm in the previous section,
it does not generate a worst case exponential overhead of ALes-formulas .
Let us now reconsider example 3.9: Given the ALes-formulas 01Fl, 01F2, and
OlF3 the optimized KD45-satisfiability algorithm constructs the ALe test sets {Fd,
27
1. Let SlF1 , . .. ,SnFn be the ACCs-formulas to be tested on KD45-satisfiability,
where each Sj is a (possibly empty) sequence of modalities and each Fj is an
A.cC- formula.
2. for i := 1, ... ,n do
if F; E Sk for some k E {I, ... , i - I }
then Sj := 0
else
Sj := {F;}
for j := i + 1, .. . ,n do
if indexes(Sj) = indexes(Sj) and
there is no position k such that Sdk] = Sdk] = O[ for some agent I
then Sj := Sj U {Fj }
endfor
endfor
3. For each non-empty set s in {Sl, ... , Sn} do: If s is not satisfiable, then STOP
and return "KD45-unsatisfiable".
4. Return "KD45-satisfiable".
Figure 3: The optimized KD45-satisfiability algorithm.
{Fd, and {F3} without applying the frame algorithm. Analogously, from D1(A = C)
and Dl(A = D) the only A.ce test set {A = C, A = D} is generated immediately from
the syntactical structure of the input A.cCs-formulas (d. example 4.1).
4.2 Restricted AC-TBoxes
In Section 2 we defined an ACC- TBox as a set of terminological axioms of the form

C = D and C =f D, where C and D are concepts. However, most of the existing ter-
minological representation and inference systems (e.g., BACK [30], CLASSIC [7], KRIS
[2]) only allow terminological axioms of the form A = C, where A is a primitive con-
cept and C is a concept. Such a terminological axiom is called (concept) definition of
A.6 Building upon this, an A.ce- TBox is then defined as a finite set of terminological
axioms which satisfies the following restrictions:
each atomic concept appears at most once as the left hand side of a terminological
axiom, and
60ften so-called concept specializations of the form A ~ C are allowed which abbreviate the
terminological axiom A = en A* where A* is a new primitive concept.
28
in this set cycles do not occur.
Thereby, a set S of terminological axioms contains a cycle iff there exists a termino-
logical axiom A = C in S such that A occurs in the concept C' which arises from C
by iterated substitutions of primitive concepts in C by the right hand sides of their
definition in S. For example, if A and B are primitive concepts the sets {A = A} and
{A = C n B, B = D u 3 R..A.} of terminological axioms contain cycles. In the following
we will call AC- TBoxes satisfying the additional conditions described above restricted
AC-TBoxes in order to distinguish them from the AC-TBoxes defined in Section 2.
It can be shown that each restricted AC- TBox 7 can be transformed into an
equivalent restricted A.cC- TBox 7' such that each right hand side of a concept def-
inition in 7' does only contain concepts which do not occur as a left hand side in
7' (see, e.g., [20]). For example, if AI, A 2, A3 are primitive concepts, the restricted
AC- TBox 7 = {AI = A2 n A3, A2 = CUD, A3 = 3 R.C} can be transformed into
7' = {AI = (C U D) n 3 R.C, A2 = CUD, A3 = 3 R.C}. Thus, each primitive concept
A on the left hand side of a terminological axiom A = C in 7' can be seen as an abbre-
viation for the concept C. With this it is easy to verify that testing consistency of an
AC-ABox A w.r.t. a restricted AC- TBox 7 is equivalent to only testing consistency
of an AC-ABox A' . Thereby, A' arises from A by successively replacing all primitive
concepts by the right hand sides of their definitions in 7. The size of A' is worst case
exponential in in the size of A and 7 (see, e.g., [20]) and testing satisfiability of A' is
known to be PSPACE-complete (see [19]). Possible optimizations are discussed in [IJ.
An algorithm for testing consistency of AC test sets which may contain termino-
logical axioms as defined in Section 2 has been given in [21J. As an easy consequence of
a result by Fischer and Ladner [13J this test is EXP- TIME complete. Moreover, when
using more expressive terminological logics than A.cC this test becomes undecidable
(for the terminological logic A.cC;: this has been shown in [4]), while this is not the
case when using restricted A.cC- TBoxes only.
Let now S be a set of ACB-formulas. Because of the above given discussion on
efficiency and decidability of testing satisfiability of an ALC- TBox and an AC-ABox it
is an interesting question whether or not the terminological axioms in each A.cC test set
which is generated from S define a restricted AC- TBox. The answer to this question
can be given with the help of Theorem 4.6 which can be used to formulate sufficient
syntactical conditions which-if satisfied-guarantee that only restricted AC- TBoxes
have to be tested in order to test KD45-satisfiability of a set of ACB-formulas. For
example, these conditions could be given by
1. agents only have positive beliefs, I.e., negation signs do not occur in front of
O-operators, and
2. for each sequence 5 of modalities holds that the set of A.cCs-formulas occurring
in the scope of 5 define a restricted AC- TBox.
29
For practical applications, however, such conditions seem not reasonable and, even
worse, when computing logical consequences (see next subsection) such syntactical
conditions in general cannot be maintained. Hence, for testing satisfiability of A.cC
test sets we in general have to take terminological axioms into account which do not
define a restricted A.cC- TBox.
4.3 Computing ACes-Inferences
We will now show how to use the KD45-satisfiability algorithm in order to test whether
or not a given ACCs-formula is a logical consequence from a set F I , . . . ,Fn of ACCB-
formulas. Again, we are only interested in KD45 Kripke structures and thus define: F
is a /(D45 consequence of FI , ... , Fn iff for each KD45 Kripke structure /( = (W, f, /(1)
and for each world w in W holds: if /(, w 1= FI, ... ,Fn, then /(, w 1= F. Firstly, let
F be an A.cCs-formula of the form o*(C = D), o*(C =/:. D), or o*(a : C), where 0* is
an abbreviation for a (possibly empty) sequence of modalities. Then, F is an KD45
consequence of FI , ... , Fn iff the set FI , ... , Fn, [-,F]* of A.cCs-formulas is not KD45-
satisfiable, where [-,F]* denotes the negation normal form of -,F. Note, that -,F is an
A.cCs-formula if F is of the above described form.
If, on the other hand, F is of the form O*(aRb), where 0* is an abbreviation for
a (possibly empty) sequence of non-negated indexed 0 operators, we cannot use this
test method since negation signs are not allowed in A.cCs-formulas which contain a
role instance. To handle this case, we extend the notion of A.cCA:-formulas as follows:
if R is a role, a, b are objects, and ill" ., im are agents, then Oil'" Oi,.,.(aRb) is an
A.cCs-formula.
Alternatively, these A.cCs-formulas could be defined by Oil' .. aim (aR'b) where (i)
each Oij is either 0ij or -,0ij' (ii) R' is either R or -,R, and (iii) the number of negation
signs in Oil' .. aim (aR' b) is even. Using this definition it is easy to see that the negation
normal form of the new A.cCs-formulas does not contain negation of roles. Therefore,
on a technical level we could allow such formulas as A.cCA:-axioms in Section 2. But a
restriction like "the number of negation signs is even" seems not to be adequate when
defining a language to describe beliefs of agents. However, for testing whether or not
an A.cCs-formula is entailed by a set FI , . . . ,Fn of Ai:Cs-axioms, this definition turns
out to be reasonable.
Note, that KD45-satisfiability of a set of A.cCs-formulas can be handled by the
KD45-satisfiability algorithm in Section 3 even if we use the above introduced extended
definition of A.cCs-formulas: Firstly, the algorithm only treats the modalities of A.cCs -
formulas, i.e., it works independently from the syntactical structure of formulas without
modalities. Secondly, satisfiability of the resulting A.cC test sets still can be tested,
since they do not contain negation of roles. And, finally, it does not matter whether aRb
is in an A.cC test set because of an input A.cCs-formula Oil'" Oim(aRb), or because
30
of an input ALCs-formula Oil .. . Oim(aRb) . Summing up, when using the extended
definition of ALCs-formulas we need not to change the KD45-satisfiability algorithm
at all. The following proposition provides a test whether or not an ALes-formula
Oil'" DiJaRb) is entailed by a set of ALes-formulas.
Proposition 4.7 Let FI , ... , Fn be a finite set of ALCs-formulas such that Fl, ... , Fn
are KD45-satisfiable, let a and b be opjects, let R be a role, and let i l , ... , im be agents.
Then Oil ... Dim (aRb) is a K D45 consequence of FI , . .. , Fn iff Oil ... Dim (aRb) lS one
of the ALCs-formulas in FI , ... , Fn
Proof: The test whether or not Oil'" Dim(aRb) is a KD45 consequence of FI , .. . , Fn is

equivalent to testing whether or not FI , ... , Fn , Oil' " Oim(a,Rb) is KD45-satisfiable.
However, since Oil' " Oim(a,Rb) is not an ALes-formula, this case cannot be handled
by the KD45-satisfiability algorithm of Section 3. Alternatively, let us have a look
at the application of the frame algorithm to the world constraint system W which is
induced by {FI , ... , Fn , Oil'" Oim(aR'b)}, where R' is a role which does not occur
in F1 , . ,Fn. Since F1 , . .. , Fn are KD45-satisfiable, each ALC test set is satisfiable
which is constructed by applying the frame algorithm to the world constraint system
which is induced by {Ft, ... ,Fn}. Hence, it is easy to verify that the ALCs-formulas
F I ,.., Fn , Oil '" Oim(aR'b) are KD45-satisfiable.
Let us now consider R' as an abbreviation for -,R. Obviously, this does not influence
the result of applying the frame algorithm to W, bU,t it may influence satisfiability of
the thereby computed ALC test sets-say A(WI),"" A(wk)-in which aR'b occurs.
It is easy to verify that, given an ALC-TBox T and an ALC-ABox A, the ALC-
formula aRb is logically entailed by T and A iff aRb E A. Hence, an ALC test
set A(Wi) E {A(Wl), ... , A(Wk)} is unsatisfiable iff aRb and aR'b are both in A(Wi).
Because of Theorem 4.6 the ALC-formulas aRb and aR'b can only be in the same
ALC test set if there is a labeled ALCs-formula Oil' .. Dim (aRb) in W. From this the
proposition follows immediately. 0
Summing up, we have now given algorithms for deciding KD45-satisfiability of a

given set of ALCs-formulas, and, building upon this, for deciding whether or not a given
ALCs-formula F is a KD45 consequence of a given set F1 , ... ,Fn of ALCs-formulas.
5 Conclusion
We have presented a two-dimensional extension of the concept language ALC which

allows both reasoning on the objective level and reasoning on the level of epistemic
alternatives. In the obtained language ALCs , a world agents are acting in can be
described by a set of terminological and assertional axioms. Furthermore, the beliefs
agents have about this world, about the beliefs of other agents, and about their own
31
beliefs can be described by terminological and assertional axioms with a leading in-
dexed 0 operator or a leading sequence of indexed 0 operators. We presented sound
and complete algorithms to check consistency of the represented beliefs and to decide
whether an AC 6 -formula is logically entailed by a given set of AC6 -formulas. Thus,
it is possible to equip agents with a decidable component to represent beliefs that is
much more expressive than representing beliefs via propositional logic. Since the ac-
tions a single agent can perform are essentially based on his local beliefs this component
can be seen as one of the basic parts in the architecture of agents.
The main restriction of the presented language AC 6 lies in the fact that modalities
are only allowed in front of terminological and assertional axioms. As an extension one
might think of modalities in front of concepts as well. Such a language would allow to
represent facts like "the things agent i believes to be expensive are exactly the things
agen t j believes to be cheap" by
Di(expensive) = Dj(cheap).
Such an extended language, however, causes algorithmic problems that are beyond the
scope of this paper and is currently investigated.
Acknowledgements. I am grateful to Hans-liirgen Biirckert, Bernhard Hollunder,
and Andreas Nonnengart for many discussions on the topic of this paper and for reading
earlier drafts.
32
References
[1] F. Baader, E. Franconi, B. Hollunder, B. Nebel, and H.-J. Profitlich. An empirical
analysis of optimization techniques for terminological representation systems or:
Making K:RIS get a move on. In Proceedings of the 3rd International Conference
on Knowledge Representation and Reasoning, Cambridge, Mass., 1992.
[2] F. Baader and B. Hollunder. K:RIS: K:nowledge Representation and Inference

System. SIGART Bulletin, 2(3):8- 14, 1991.
[3] F. Baader and B. Hollunder. A terminological knowledge representation system

with complete inference algorithms. In M. Richter and H. Boley, editors, Interna-
tional Workshop on Processing Declarative Knowledge, 1991.
[4] F. Baader and B. Hollunder. Embedding defaults into terminological knowledge

representation formalisms. In Proceedings of th e 3rd International Conference on
Knowledg e Representation and Reasoning, Cambridge, Mass., 1992.
[5] F . Baader and H.-J. Ohlbach. A multi-dimensional terminological knowledge rep-

resentation language. In Proceedings of IlCAI 93, 1993. To appear.
[6] A. Bond and L. Gasser. R eadings in Distribut ed Artificial Intelligence. Morgan

Kaufmann, Los Angeles, CA, 1988.
[7] A. Borgida, R. J. Brachman, D. 1. McGuinnes~, and 1. A. Resnick. CLASSIC:
a structural data model for objects . In Proceedings of the 1989 ACM SIGMOD
International Conference on Managem ent of Data, pages 59- 67, Portland, Oreg.,
1989.
[8] R. Brachman and H. J . Levesque. Expressiveness and tractability in knowledge

representation and reasoning. Computational Intelligence, 3:78-93, 1987.
[9] R. J. Brachman and J. G. Schmolze. An overwiev of the KL-ONE knowledge

representation system. Cognitive Science, 9(2):171-216, 1985.
[10] B. F. Chellas. Modal Logic: An Introduction. Cambridge University Press, 1980.
[11] N. Dershowitz and Z. Manna. Proving termination with multiset orderings . ACM,
22(8):465- 476, 1979.
[12] R. Fagin and J. Y. Halpern. Two views of belief. Artificial Intelligence, 54(3):275-
317., April 1992.
[13] M. J. Fischer and R. E. Ladner. Propositional dynamic logic of regular programs.

Journal of Computer and System Science, 18:194-211, 1979.
[14] M. Fitting. Proof Methods for Modal and Intuitionistic Logics, volume 169 of
Synthese Library. D. Reidel Publishing Company, 1983.
33
[15] 1. Gasser and M.N. Huhns. Distributed A rtificial Intelligence, Volume II. Research
Notes in Artificial Intelligence. Morgan Kaufmann, San Mateo, CA, 1989.
[16] J. Y. Halpern and Y. Moses. Knowledge and common knowledge in a distributed

environment. Journal of the ACM, 37(3):549-587, 1990.
[17] J. Y. Halpern and Y. Moses. A guide to completeness and complexity for modal
logics of knowledge and belief. Artificial Intelligence, 54 :319- 379, 1992.
[18] J. Hintikka, editor. Knowledge and Belief Cornell University Press, 1962.
[19] B. Hollunder. How to reduce reasoning to satisfiability checking of concepts in the

terminological system KRIS. To appear.
[20] B. Hollunder. Hybrid inferences in KL-ONE-based knowledge representation

systems. In 14th German Workshop on Artificial Intelligenc e, volume 251 of
Informatik-Fachberichte, pages 38-47, Ebingerfeld, Germany, 1990. Springer.
[21] A. Laux. Integrating a modal logic of knowledge into terminological logics. Re-
search Report RR-92-56, DFKI Saarbriicken, 1992.
[22] W. Lenzen. Recent work in epistemic logic. Acta Philosophica Fennica, 30:1- 219,
1978.
[23] H. J. Levesque. Foundations of a functional approach to knowledge representation.

Artificial Intelligence, 23:155-212, 1984.
[24] G. 1. McArthur. Reasoning about knowledge and belief: A survey. Computational

Intelligence, 4:223-243, 1988.
[25] J.-J. C. Meyer, W. van der Hoek, and G. A. W. Vreeswijk. Epistemic logi c for
computer science: A tutorial (part one). In Bulletin of the EATCS, volume 44,
pages 242-270. European Association for Theoretical Computer Science, 1991.
[26] J.-J. C. Meyer, W. van der Hoek, and G. A. W. Vreeswijk. Epistemic logic for
computer science: A tutorial (part two). In Bulletin of the EATCS, volume 45,
pages 256-287. European Association for Theoretical Computer Science, 1991.
[27] R. C. Moore. Reasoning about knowledge and action. Technical Report 191, SRI
International, 1980.
[28] R. C. Moore. A formal theory of knowledge and action. In J. R. Hobbs and

R. C. Moore, editors, Formal Theories of the Commonsense World, pages 319-
358. Ablex Publishing Corporation, 1985.
[29] B. Nebel. Reasoning and Revision in Hybrid Knowledge Representation Systems.

Number 422 in LNAI. Springer, 1990.
34
[30] B. Nebel and K. von Luck. Hybrid reasoning in BACK. In Z. W. Ras and
L. Saitta, editors, Methodologies for Intelligent Systems, volume 3, pages 260-269.
North-Holland, 1988.
[31] K. Schild. A correspondence theory for terminological logics: Preliminary report.

In Proceedings of IJeAI 91, pages 466-471, 1991.
35
Deutsches DFKI
Forschungszentrum -Bibliothek-
fOr KOnstliche PF 2080
Intelligenz GmbH 67608 Kaiserslautern
FRG
DFKI Publikationen DFKI Publications

Die folgenden DFKI VerOffentlichungen sowie die The following DFKI publications or the list of all
aktuelle Liste von allen bisher erschienenen published papers so far can be ordered from the
Publikationen konnen von der oben angegebenen above address.
Adresse bezogen werden. The reports are distributed free of charge except if
Die Berichte werden, wenn nicht anders otherwise indicated.
gekennzeichnet, kostenlos abgegcben.
DFKI Research Reports RR9250

Stephan Busemann:
RR9242 Generierung natiirlicher Sprache
John Nerbonne: 61 Seiten
A Feature-Based Syntax/Semantics Interface
19 pages RR9251
!-!ans-Jurgen Burcken. Werner Nutt:
RR9243 On Abduction and Answer Generation through
Christoph Klauck. Jakob Mauss : A Heuristic driven Constrained Resolution
Parser for Attributed Node Labeled Graph 20 pages
Grammars and its Application to Feature
Recognition in CIM RR9252
17 pages Mathias Bauer. Susanne Biundo. Dietmar Dengler.
Jana Koehler. Gabriele Paul: PHI - A Logic-Based
RR9244 Tool for Intelligent Help Systems
Thomas Rist. Elisabeth Andre: Incorporating 14 pages
Graphics Design and Realization into the
Multimodal Presentation System WIP RR9253
15 pages Werner Stephan. Susanne Biundo:
A New Logical Framework for Deductive Planning
RR9245 15 pages
Elisabeth Andre. Thomas Rist: The Design of
Illustrated Documents as a Planning Task RR9254
21 pages I-Iarold Boley: A Direkt Semantic Characterization
ofRELFUN
RR9246 30 pages
Elisabeth Andre. Wolfgang Finkler. Win/ried Gra/.
Thomas Rist. Anne Schauder. Wolfgang Wahlster : RR9255
WIP: The Automatic Synthesis of Multimodal John Nerbonne. Joachim Laubsch. Abdel Kader
Presentations Diagne. Stephan Oepen: Natural Language
19 pages Semantics and Compiler Technology
17 pages
RR9247
FrankBomarius: A Multi-Agent Approach towards RR9256
Modeling Urban Traffic Scenarios Armin Lata: Integrating a Modal Logic of
24 pages Knowledge into Terminological Logics
34 pages
RR9248
Bernhard Nebel. Jana Koehler: RR9258
Plan Modifications versus Plan Generation: Franz Baader. Bernhard Hol/under:
A Complexity-Theoretic Perspective How to Prefer More Specific Defaults in
15 pages Terminological Default Logic
31 pages
RR9249
Christoph Klauck. Ralf Legleitner. Ansgar Bernardi: RR9259
Heuristic Classification for Automated CAPP Karl Schlechta and David Makinson: On Principles
15 pages and Problems of Defeasible Inheritance
13 pages
RR-92-60 RR-93-13
Karl Schlechta: Defaults, Prcorder Semantics and Franz Baader, Karl Schlechta :
Circumsc'ription A Semantics for Open Normal Defaults via a
19 pages Modified Preferential Approach
25 pages
RR-93-02
Wolfgang Wahl.~ter, Elisabeth Andre, Wolfgang RR-93-14
Finkler,lIans-Jurgen Profit/ich, Thomas Rist: Joachim Niehren, Andreas Podelski.RaIfTreinen:
Plan-based Integration of Natural Language and Equational and Membership Constraints for
Graphics Generation Infinite Trees
50 pages 33 pages
RR-93-03 RR-93-1S
Franz Baader, Berhard lIollunder, Bernhard Frank Berger, Thomas Fehrle, Kristof KlOckner,
Nebel,lIans-Jurgen Profit/ich, Enrico Franconi: Volker SchOlles, Markus A. Thies, Wolfgang
An Empirical Analysis of Optimization Techniques Wahlster : PLUS - Plan-based User Support
for Terminological RepresenLation Systems Final Project Report
28 pages 33 pages
RR-93-04 RR-93-16
Christoph Klauck, Johannes Schwagereit: Gert Smolka, Martin Ifenz , Jorg Wurtz: Object-
GGD: Graph Grammar Developer for features in Oriented Concurrent Constraint Programming in
CAD/CAM 01.
13 pages 17 pages
RR-93-05 RR-93-17
Franz Baader, Klaus Schulz : Combination Tech- Rolf Backofen:
niques and Decision Problems for Disunification Regular Path Expressions in Feature Logic
29 pages 37 pages
RR-93-06 RR-93-1S
Hans-Jurgen Bi1.rckert, Bernhard Ifollunder, Armin Klaus Schild: Tenninological Cycles and the
Laux: On Skolemization in Constrained Logics Propositional j.l-Calculus
40 pages 32 pages
RR-93-07 RR-93-20
Hans-Jurgen Burckert, Bernhard Ifollunder, Armin Franz Baader, Bernhard Ifollunder:
Laux: Concept Logics with Function Symbols Embedding Defaults into Tenninological
36 pages Knowledge RepresenLation Formalisms
34 pages
RR-93-08
Harold Boley, Philipp Hanschke, Knutllinkelmann, RR-93-22
Manfred Meyer: COLAB: A Hybrid Knowledge Manfred Meyer, Jorg Muller:
RepresenLation and Compilation Laboratory Weak Looking-Ahead and its Application in
64 pages Computer-Aided Process Planning
17 pages
RR-93-09
Philipp Hanschke, Jorg Wurtz : RR-93-23
Satisfiability of the Smallest Binary Program Andreas Dengel, Oltmar Lutzy :
8 Seiten Comparative Study of Connectionist Simulators
20 pages
RR-93-10
Martin Buchheit, Francesco M. Donini, Andrea RR-93-24
Schaerf: Decidable Reasoning in Terminological Rainer Ifoch, Andreas Dengel:
Knowledge RepresenLation Systems Document Highlighting -
35 pages Message Classification in Printed Business Letters
17 pages
RR-93-11
Bernhard Nebel, Hans-Juergen Buerckert: RR-93-25
Reasoning about Temporal Relations: Klaus Fischer, Norbert Kuhn: A DAI Approach to
A Maximal TracLable Subclass of Allen's Interval Modeling the TransporLation Domain
Algebra 93 pages
28 pages
RR-93-26
RR-93-12 Jorg P. Muller, Markus Pischel : The Agent
Pierre Sablayrolles: A Two-Level Semantics for Architecture InteRRaP: Concept and Application
French Expressions of Motion 99 pages
51 pages
RR-93-27 01'1'.1 Technical Memos
Hans-Ulrich Krieger:
Derivation Without Lexical Rules TM-91-15
33 pages Stcfan Busemann: Prototypical Concept Formation
An Alternative Approach to Knowledge Representation
RR-93-28 28 pages
/lans-Ulrich Krieger, lohn Nerbonnc,
Hannes Pirker: FcaLUrc-Bascd Allomorphy TM-92-01
8 pages Lijuan Zhang: Entwurf und Implcmentierung eines
Compilers zur Transformation von
RR-93-29 Werkstlickreprasentationen
Armin Laux: Representing Belief in Multi -Agent 34 Seiten
Worlds viaTerminological Logics
35 pages TM-92-02
Achim Schupeta: Organizing Communication and
RR-93-33 Introspection in a Multi-Agent Blocksworld
Bernhard Nebel, lana Koehler: 32 pages
Plan Reuse versus Plan Generation: A Theoretical
and Empirical Analysis TM-92-03
33 pages Mona Singh:
A Cognitiv Analysis of Event Structure
RR-93-34 21 pages
Wolfgang Wahlster:
Verbmobil Translation of Face-To-Face Dialogs TM-92-04
10 pages lurgen Muller , lorg Muller, Markus Pischel,
Ralf Scheidhauer:
RR-93-35 On the Representation of Temporal Knowledge
Harold Boley, Franc;ois Bry, Ulrich Geske (Eds .): 61 pages
Neuere Entwicklungen der deklarativen KI -
Programmierung - Proceedings TM-92-0S
150 Seiten Franz Schmalhofer, Christoph Globig, lorg Thoben:
Note: This document is available only for a The refitting of plans by a human expert
nominal charge of25 DM (or 15 US-S) . 10 pages
RR-93-36 TM-92-06
Michael M. Richter, Bernd Bachmann , Ansgar Gila Kuhn , Franz Schmalhofer: Hierarchical
Bernardi, Christoph Klau ck, Ralf Legleitner , skeletal plan refinement: Task- and inference
Gabriele Schmidt: Von IDA bis IMCOD: structures
Expertensysteme im CIM-Umfeld 14 pages
13 Seiten
TM-92-08
RR-93-38 Anne Kilger: Realization of Tree Adjoining
Stephan Baumann : Document Recognition of Grammars with Unification
Printed Scores and Transformation into MIDI 27 pages
24 pages
TM-93-01
RR-93-40 Gila Kuhn, Andreas Birk: Reconstructive
Francesco M. Donini , Maurizio Lenzerini, Daniele Integrated Explanation of Lathe Production Plans
Nardi, Werner NUll, Andrea Schaerf: 20 pages
Queries, Rules and Definitions as Epistemic
Statements in Concept Languages TM-93-02
23 pages Pierre Sablayrolles, Achim Schupeta:
Conlfict Resolving Negotiation for COoperative
RR-93-41 Sc hedule Management
WinJried H. Graf: LA YLAB : A Constraint-Based 21 pages
Layout Manager for Multimedia Presentations
9 pages TM-93-03
Harold Boley, Ulrich Buhrmann, ChristoJ Kremer:
RR-93-42 Konzeption einer deklarativen Wissensbasis tiber
Hubert Comon, RalfTreinen : recyclingrelevante Materialien
The First-Order Theory of Lexicographic Path II pages
Orderings is Undecidable
9 pages TM-93-04
Hans-Gtinther Hein: Propagation Techniques in
RR-93-4S W AM-based Architectures - The FlOO-III
Rainer Hoch: On Virtual Partitioning of Large Approach
Dictionaries for Contextual Post-Processing to lOS pages
Improve Character Recognition
21 pages
OFKI Oocuments 09304
DFKI Wissenschaftlich-Technischer Jahresbericht
09219 1992
Stefan Dittrich. Rainer Hoch: Automatische, 194 Seiten
Deskriptorbasierte Unterstiitzung der Dokument-
analyse zur Fokussierung und Klassifizierung von 09305
Geschaftsbriefen Elisabeth Andre. Winfried Graf. lochen Heinsohn.
107 Seiten Bernhard Nebel. Hans-lurgen Profitlich. Thomas
Rist . Wolfgang Wahlster :
09221 PPP: Personalized Plan-Based Presenter
Anne Schauder: Incremental Syntactic Generation 70 pages
of Natural Language with Tree Adjoining
Grammars 09306
57 pages lurgen M wler (Hrsg.):
Beitrage zum Griindungsworkshop der Fachgruppe
09222 Verteilte Kiinstliche Intelligenz Saarbriicken 29.-
Werner Stein: Indexing Principles fer Relational 30. April 1993
Languages Applied to PROLOG Code Generation 235 Seiten
80 pages Note: This document is available only for a
nominal charge of 25 DM (or 15 US-$).
09223
Michael Herfert : Parsen und Generieren der 09307
Prolog-artigen Syntax von RELFUN Klaus-Peter Gores. Rainer Bleisinger:
51 Seiten Ein erwartungsgesteuerter Koordinator zur
partiellen Textanalyse
09224 53 Scitcn
lurgen Muller. Donald Steiner (f1rsg.):
Kooperierende Agenten 09308
78 Seiten Thomas Kieninger. Rainer Hoch : Ein Generator
mit Anfragesystem fiir strukturierte Worterbiicher
09225 zur Untcrstiitzung von Texterkcnnung und
Martin Buchheit: Klassische Kommunikations- und Textanalyse
Koordinationsmodclle 125 Seiten
31 Seiten
09309
09226 Han s-Ulrich Krieger. Ulrich Schafer:
Enno Tolzmann: TDL ExtraLight User's Guide
Realisierung eines Werkzeugauswahlmoduls mit 35 pages
Hilfe des Constraint-Systems CONTAX
28 Seiten D9310
Elizabeth lIinkelman . Markus Vonerden.Christoph
09227 lung : Natural Language Software Registry
Martin Harm. Knut Hinkelmann. Thomas Labisch: (Second Edition)
Integrating Top-down and BOllom-up Reasoning in 174 pages
COLAB
40 pages 0-9311
Knut llinkelmann. Armin Laux (Eds.):
09228 DFKI Workshop on Knowledge Representation
Klaus-Peter Gores. Rainer Bleisinger: Ein Modell Techniques - Proceedings
zur Reprasentation von Nachrichtentypen 88 pages
56 Seiten
D9312
09301 Ilaroid Boley. Klaus Elsbernd. Michael Herfert.
Philipp Hanschke. Thom Fruhwirth : Termi nological Michael Sintek. Werner Stein :
Reasoning with Constraint Handling Rules RELFUN Guide: Programming with Relations and
12 pages Functions Made Easy
86 pages
09302
Gabriele Schmidt. Frank Peters. 09314
Gernod Laufkotter: User Manual of COKAM+ Manfred Meyer (Ed.) : Constraint Processing-
23 pages Proceedings of the International Workshop at
CSAM'93, July 20-21,1993
09303 264 pages
Stephan Busemann. Karin Ilarbus ch( Eds.): Note: This document is available only for a
DFKI Workshop on Natural Language Systems: nominal charge of 25 DM (or 15 US-$).
Reusability and Modularity - Proceedings
74 pages
.,l> JJ
(I)
~. "C
::;,
'"'
(I)
I' en
-
I\) (I)
c :l
)(
:l
to
OJ
(I)
-
(I)
:::J
3:
c::
;:;
I

to
-:e
(I)
:::J
0
'"'
Q.
en
<
D)
-4
(I)
3'"'
_.
:l
2-
0
to
(')
D)
r
0
to
(')
en

Multi-Agent Belief

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Multi-Agent Belief

Uploaded by

Copyright:

Available Formats

Deutsches

Representing Belief in Multi-Agent Worlds

Deutsches Forschungszentrum fur Kunstliche Intelligenz

o Intelligent Engineering Systems

Deutsches Forschungszentrum fUr Kunstliche Intelligenz 1993

2 Syntax and Semantics of ACB 5

3 Testing Satisfiability of ACB-formulas 9

4 Optimization and Computing ACB-Inferences 20

respectively, where Di<.p is to be read as "agent i believes <.p"

2 Syntax and Semantics of ALeE

2.1 The Concept Language ALe

Terminological logics provide two formalisms to describe a problem domain: a termino-

1. Each atomic concept, T, and 1- are concepts.

2. If e and D are concepts and R is a role, then

(a) enD (concept conjunction),

An interpretation I is a function over some non-empty domain L\ I which maps each

[VR.CjI = {d E L\I I Vd': (d,d') E RI -+ d' Eel}

e i- D (negated concept equivalence)

carrier = person n 3 owns. truck.

The assertional formalism of AC allows one to introduce concrete objects by stat-

John owns truck-l and truck-l : truck.

Thus, we can describe the relevant concepts of a problem domain by terminological

LelTIlTIa 2.1 Let C and D be concepts, and let I be an interpretation. Then

1. I satisfies C ~ D iff I satisfies -,C U D = T.

2. I satisfies C (l D iff I satisfies -,C U D -::J T.

2.2 The Extended Language ALes

If TAis a terminological axiom, then OJ TA and --.ojTA are terminological axioms

If CJ is a concept instance, then OJ CJ and --. OJ CJ are concept instances as well.

If RI is a role instance, then OJ RI is a role instance as well.

The satisfiability of an ALCs -formula F in a /(ripke structure /( = (W, r,I<[) and a

/(, w F= C = D iff {d I (d,w) E CKI} = {d I (d,w) E DKI}

where G is an ALCs-formula, C , D are concepts, a , b are objects, and R is a role.

3 Testing Satisfiability of ALes-formulas

3.1 The A.ces Frame Algorithm

serial iff for each u in W there is a v in W such that (u, v) E "

Euclideaniffforallu,v,win Wholds: if(u,v) E ,and (u,w) E ,then (v,w) E 1,

(PI') if K, w 1= DjF then K, w 1= 0 F j

"F -t F ,(C = D) -t an: (C n ,D) U (,C n D)

to the outermost negation sign. Thereby, F is an ACB-formula, C, D are concepts,

Lemma 3.2 Let F be an ACB-formula, F* be the negation normal form of F, and

If F is an ACB-formula in negation normal form it has a (possibly empty) leading

constraint system is a finite, non-empty set of world constraints.

W --+0 {w ~i v, G l II v, DiG l II v, ... , Gm II v, OiGm II v} U W

Figure 1: Propagation rules of the ACCB frame algorithm.

guarantee termination of applying propagation rules to a derived system.

Proposition 3.3 If W is a world constraint system which is induced by a finite set of

5. The mapping 1f1w (w) is now defined by the pair

of integers, and W(W) is the multiset {t/;w (w) Iw is a label in W}.

{F I F II w E W' and F does not contain any modality}.

where v is a new label. By induction hypothesis there is a world w K in W such that

Proof: Let K be the Kripke structure (W, f, K 1 ) where

W is given by the set of all labels in W'.

f consists of an accessibility relation Ii for each agent i. Thereby, Ii is given by

KI is given such that K , w 1= F for each labeled A.cCB-formula F II w in W'

2. there are two paths starting with some label z, namely

(*) if DiF II w is in W', then DjF II z is also in W'.

It is sufficient to show that DiF II u is in Wj whenever DiF II u' is in Wj + b where Wj - 0

2. Let W' be the result of the frame algorithm with input W.

Figure 2: The KD45-satisfiability algorithm.

The next theorem summarizes the previous results.

Proof; The ACs-formulas F 1 , . , Fn are KD45-satisfiable iff W is KD45-satisfiable.

Summing up, we obtain the algorithm for testing KD45-satisfiability of AC s -

WI Wo U {wo ~l WI, Fl II w}, 01F211 WI,, OlFn II wd

oOjj o~ F /I v, oOj) 0; G /I v, OJ 0* Oij 0; G /I v (a)

OJ 0* OJ) o~ F /I v, OOi) o~ F /I v, OJ 0 Ojj 0; G /I v, O*Oij 0; G /I v (J)

(1) oOjj o~ F II v and oOjj 0; G II v are in WI or